Speed up EM hot paths without changing results

[olivierC80]

Summary

This PR reduces allocation and native-call overhead in poLCA while preserving
the package's results exactly for the tested execution paths.

The implementation deliberately excludes algebraically equivalent changes that
altered floating-point results in the last bits. It also excludes the
poLCA.compress() optimization because that function is already being changed
in #28.

Changes

• Compute as.integer(t(y)) once per fit and reuse it in native wrappers.
• Add PACKAGE = "poLCA" to the existing .C() calls.
• Fuse the existing postclass and probhat loops into one internal .Call()
  for models with and without covariates.
• Keep separate complete-data and missing-data native paths without changing
  the order of floating-point operations.
• Implement rmulti() with .Call() and R's RNG API while preserving both its
  numeric return type and RNG stream.
• Preallocate known-size objects in poLCA.se() without replacing or
  reordering its matrix calculations.
• Detect constant manifest columns directly.
• Use max.col(..., ties.method = "first") for modal assignment.

Exact-result validation

The candidate was compared with commit
2ffbf2fd44fe707c54e3939fe929b8032b5db63f using separately installed source
packages. No rounding or numeric tolerance was used: results were checked with
identical() after removing only the nondeterministic elapsed-time field.

The strict comparison covered:

• fits with and without standard errors;
• models with and without covariates;
• complete and missing manifest data;
• one-, three-, and four-class models;
• repeated estimation with nrep = 3;
• constant manifest variables;
• posterior probabilities, predicted cells, tables, entropy, coefficients,
  and variance-covariance matrices;
• matrix and vector inputs to rmulti(), including the resulting RNG state.

All compared objects were strictly identical, including types, dimensions,
attributes, and floating-point values.

The new package test additionally compares the fused native helpers directly
with the historical postclass and probhat wrappers for shared/varying priors
and complete/missing data. It also checks exact rmulti() output and RNG state.

Performance

Mean elapsed time over 10 repetitions, original versus candidate installed
source packages:

Scenario	Original	Candidate	Speedup
fit_basic	0.0222 s	0.0161 s	1.38x
fit_covariates	0.0253 s	0.0207 s	1.22x
fit_covariates_large	0.0783 s	0.0694 s	1.13x
fit_em_large	0.3207 s	0.1765 s	1.82x
fit_em_xlarge	0.8046 s	0.4302 s	1.87x
fit_heavy	0.1099 s	0.0579 s	1.90x
fit_heavy_se	0.1394 s	0.0896 s	1.56x
fit_missing_keep_na	0.0059 s	0.0036 s	1.64x
rmulti_large	0.4100 s	0.0778 s	5.27x
simdata	0.0056 s	0.0015 s	3.73x

Package check

R CMD check --no-manual --no-build-vignettes completes with no errors or
warnings. It reports the same existing NOTE as the unmodified baseline:

Unknown package 'flexmix' in Rd xrefs

Deliberate exclusions

The following previously explored changes are not included because they did
not satisfy strict bit-for-bit equality:

• replacing t(s) %*% s with crossprod(s);
• replacing covariance products with tcrossprod();
• computing Jac.mix through an algebraically equivalent reordered formula;
• fusing log-likelihood or prior accumulation into the native EM helper.