feat(optim): add auditable parameter-group coverage

[PinkTulips139]

Summary

Adds a read-only optimizer parameter-group audit for validating how model parameters are assigned to optimizer.param_groups.

The audit makes optimizer construction more observable by checking trainable-parameter coverage, duplicate assignments, frozen parameters present in the optimizer, and optimizer parameters that cannot be mapped back to the model.

Changes

• Add audit_optimizer_param_groups(model, optimizer, strict=False).
• Add the structured OptimizerGroupAuditError.
• Report:
  • missing trainable parameters;
  • parameters assigned more than once;
  • frozen parameters present in optimizer groups;
  • optimizer parameters not owned by the model;
  • exactly-once trainable coverage;
  • tensor and element counts per group;
  • current lr, optional initial_lr, weight_decay, and explicit group metadata.
• Preserve explicit optimizer group semantics such as group_name, name, and role.
• Avoid guessing Adapter, Router, or Head roles from parameter-name substrings.
• Integrate the read-only audit after adapter/PEFT optimizer configuration and before scheduler construction.
• Store the resulting structured snapshot on the Trainer and emit a compact main-rank summary.

Strictness

With strict=False, all findings are returned as structured data without interrupting training.

With strict=True, missing trainable parameters, duplicate assignments, and unknown optimizer parameters raise OptimizerGroupAuditError.

Frozen parameters in optimizer groups are reported but are not automatically removed.

Behavior guarantees

This PR does not change:

• optimizer group contents or order;
• learning rates or learning-rate multipliers;
• initial_lr or weight decay;
• scheduler or warmup behavior;
• checkpoint or resume topology;
• model outputs;
• LoRA, Adapter, or Router defaults.

The audit is a read-only snapshot of the current optimizer state.

Validation

• 20 passed in optimizer-group audit tests
• 9 passed in PEFT optimizer-policy tests
• 3 passed in LoRA layer-decay tests
• Ruff passed
• Ruff format check passed
• Python compilation passed
• git diff --check passed

The tests cover exactly-once coverage, missing and duplicate parameters, frozen and unknown parameters, strict versus reporting modes, JSON serialization, group statistics, scheduler-updated learning rates, and non-PEFT optimizers.

[PinkTulips139]

CI note: the only failing job is the legacy Python 3.8 / PyTorch 1.8.0 environment, which fails in the existing Mixture tests because torch.autocast is unavailable. This compatibility issue is addressed separately by #185. PR #186 changes only optimizer parameter-group auditing and does not modify the shared autocast helper or Mixture computation.