fix: resolve 8 bugs in core model and inference code

[sjhddh]

Summary

Systematic audit of model/module.py and model/kronos.py uncovered 8 bugs, several with crash potential:

• requires_grad typo in FixedEmbedding — w.require_grad = False is a silent no-op (should be requires_grad)
• Gradient count mismatch in DifferentiableEntropyFunction.backward — returns 5 values for 4 forward args
• int() unpacking crash in get_codebook_entry / get_group_codebook_entry — h, w = int(...) can't unpack a scalar
• NameError when soft_entropy=False — avg_prob referenced but never assigned in that branch
• is_causal=True in cross-attention — semantically wrong (cross-attn should attend to all keys) and crashes when attn_mask is also provided
• None comparison crash in sample_from_logits — top_k > 0 raises TypeError when top_k is None
• In-place logit mutation in top_k_top_p_filtering — caller's tensor silently modified
• Missing @torch.no_grad() on KronosPredictor.generate/predict/predict_batch

Also:

• Replaced wildcard from model.module import * with explicit imports
• Fixed self-referential docstrings on indexes_to_codes / group_indexes_to_codes
• Removed commented-out code blocks
• Extracted shared _validate_and_prepare helper to deduplicate DataFrame validation

Test plan

• Existing regression tests still pass (pytest tests/test_kronos_regression.py)
• FixedEmbedding weights are now truly frozen (verify requires_grad == False)
• sample_from_logits(logits, top_k=None, top_p=None) no longer crashes
• top_k_top_p_filtering does not mutate the input tensor

🤖 Generated with Claude Code