Feat (llm/eval): Standard deviation to perplexity computation

[pablomlago]

Reason for this PR

compute_perplexity reports a single point estimate, making it hard to tell whether perplexity differences between configurations (float vs. quantized, recipe A vs. B) are meaningful. Other ecosystem tools (lm-eval-harness, lighteval) report a bootstrap stderr alongside perplexity for this reason.

Changes Made in this PR

• src/brevitas_examples/llm/llm_quant/eval.py
  • compute_perplexity now returns Tuple[float, float]: (ppl, ppl_std).
  • Stderr computed via scipy.stats.bootstrap over the per-sequence NLLs with statistic exp(mean(x)), n_resamples=1000, method="BCa".
• src/brevitas_examples/llm/main.py
  • Both call sites unpack the tuple and print "<X> perplexity (<dataset>): <ppl:.3f> ± <std:.3f>".
• src/brevitas_examples/llm/benchmark/llm_benchmark.py
  • eval_metrics extended with float_ppl_std and quant_ppl_std.
  • parse_log regex updated to capture the new ± <std> suffix as an optional group, keeping backward compatibility with older logs.

Bootstrap was chosen over a closed-form stderr because perplexity (exp(mean(nll))) is non-linear in the NLLs; this matches the lighteval/lm-eval convention.

Testing Summary

• Sanity import of the new compute_perplexity signature.
• Manual float + quantized eval run confirming the new print format and that parse_log extracts both ppl and std.
• Verified parse_log backward compatibility on logs without the ± <std> suffix (returns None for _std fields).

Risk Highlight

• This PR includes code from another work (please detail).
  • Bootstrap design follows lighteval's bootstrap_stderr_scipy (MIT, EleutherAI / HuggingFace 2024).
• This PR contains API-breaking changes.
  • compute_perplexity return type changes from float to Tuple[float, float]. In-tree call sites are updated.
• This PR depends on work in another PR (please provide links/details).
• This PR introduces new dependencies (please detail).
  • First direct use of scipy.stats.bootstrap in this module (scipy is already a transitive dep).
• There are coverage gaps not covered by tests.
  • No dedicated unit test for the bootstrap stderr; covered indirectly via existing eval/benchmark paths.
• Documentation updates required in subsequent PR.

Checklist

• Code comments added to any hard-to-understand areas, if applicable.
• Changes generate no new warnings.
• Updated any relevant tests, if applicable.
• No conflicts with destination dev branch.
• I reviewed my own code changes.
• Initial CI/CD passing.
• 1+ reviews given, and any review issues addressed and approved.
• Post-review full CI/CD passing.