Enable more clippy lints - #24322

The `.peekable()` in the sqllogictest Postgres engine was never peeked,
so it only added an extra layer of indirection.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Catches `From` implementations that can panic, where `TryFrom` would be
the honest signature.

All three existing hits would need a breaking API change to fix, so they
get `#[expect]` for now. Two of them (`Constraint`) panic on a protobuf
message with an unset `constraint_mode`, i.e. on malformed input.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Catches exact float comparisons against constants, e.g. `x == 0.0`.

The two existing hits in `value_transition!` really do want an exact
comparison against `f32::MIN`/`MAX`, so they get `#[expect]`.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Catches float literals that silently round, e.g. `let x: f32 = 0.1234567890123;`.

The three existing hits are false positives: they spell out exact powers
of two (2^64 and 2^64-2^41), which float `Display` renders with fewer
digits, so the lint thinks precision was lost.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

`(a + b) / 2` overflows when `a + b` exceeds the type's range;
`a.midpoint(b)` does not.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Catches strings that contain `{...}` but are never actually formatted.

This found two real bugs where the placeholder was silently printed
verbatim:
* `benchmarks/src/nlj.rs`: `"NLJ benchmark Q{query_id} failed…".to_string()`
* `parquet_advanced_index.rs`: `.expect("metadata for file not found: {filename}")`

The remaining hits are intentional: shell-style `${VAR:-default}`
placeholders, `{rows}`-style templates substituted with `str::replace`,
and braces inside expected struct output. Those get `#[expect]`.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

`&x as *const T` silently picks a pointer type; `std::ptr::from_ref(&x)`
keeps the referent type explicit and cannot accidentally change it.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Catches `// SAFETY:` comments that do not sit in front of anything
unsafe, so that a `SAFETY:` comment reliably means "an unsafe block
follows, and here is why it is sound".

* Three comments documented an `unwrap` or a safe copy rather than
  unsafe code, so they lose the `SAFETY:` prefix.
* Three sat in front of an `if` while the `unsafe` block was inside it,
  so they move next to the block they justify.
* Two were prose false positives, where the lint matched "safety:" in
  the middle of a doc comment.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Catches large types passed by value, which forces a memcpy at every call.

The single existing hit is `HyperLogLog::new_with_registers`, whose 16 KiB
array is moved into the returned struct, so a reference would only add a
copy. It gets `#[expect]`.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

Catches `-> Box<T>` where the caller gains nothing from the indirection.

The single existing hit is a test helper that both takes and returns
`Box<Expr>` so it can hand back the same allocation, so it gets
`#[expect]`.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

A `..` in a pattern that already binds every field does nothing today,
but silently swallows any field added later. Removing it turns that into
a compile error.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

`if let Some(true) = x` reads as a binding but is really an equality
check; `x == Some(true)` (or `matches!`) says so.

Clippy suggested one tuple comparison, `(is_valid, is_included) == (true,
Some(true))`; that one is written as a plain `&&` instead.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

`x.deref()` and `x.deref_mut()` are the operator spelled the long way;
`&*x` / `&mut *x` is the idiomatic form and does not need `Deref` in
scope.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

This reverts commit c7e25ff.

Addresses review feedback on the `equatable_if_let` commit: clippy
suggested `matches!` in a few places where a plain equality check reads
better.

`predicate_bounds.rs` uses `.ok() == Some(false)` because
`DataFusionError` does not implement `PartialEq`, so `== Ok(false)` does
not compile.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

This reverts commit d82dc41.

This reverts commit 36abeb7.

This reverts commit 85c06c6.

* `test_parse_duration_with_overflow_check` uses `Duration::from_mins`
  for the `"…m"` input again, so the constructor mirrors the unit suffix
  in the string being parsed. That trips
  `duration_suboptimal_units`, so the test gets an `#[expect]` saying why.
* Restore the `TODO`/`Issue` comments above the ignored
  `sort_with_mem_limit_2_cols_2` test, keeping a short `#[ignore]` reason.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

The upstream merge brought in code written before these lints were
enabled, so it failed `clippy -D warnings` on the merge commit:

- `branches_sharing_code`: hoist `raw_keys` out of the if/else in
  `dictionary.rs`
- `iter_on_single_items`: `[None].iter()` -> `std::iter::once(&None)`
- `elidable_lifetime_names`: elide `'a` in `scan_with_args_inner`,
  `infer_boxed`, and `infer_options_boxed`

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

…ilk/more-clippy-lints