fix(qwen4): freeze forward-only QSA indexer

[leooop-al]

Summary

Freeze QSAIndexer parameters after construction so they are excluded from
optimizer groups and Megatron DDP gradient buckets.

Problem

All QSA selection paths run under torch.no_grad() and produce discrete boolean
masks or integer token indices. There is currently no auxiliary QSA indexer loss,
so index_qk_proj, q_layernorm, and k_layernorm cannot receive gradients from
the language-model loss.

Despite that, these parameters default to requires_grad=True. Megatron DDP then
adds them to gradient buckets and installs backward hooks that never fire. With
overlapped gradient reduction enabled, the next zero_grad_buffer() fails because
the bucket ready count is incomplete.

A two-node reproduction reported exactly these missing parameters:

first-batch hook mismatch: ready=31 expected=34 missing=3
indexer.index_qk_proj.weight
indexer.k_layernorm.weight
indexer.q_layernorm.weight

Fix

Call self.requires_grad_(False) after all QSAIndexer parameters are created.
This preserves checkpoint loading and forward selection while keeping the
forward-only parameters out of optimizer state and DDP gradient buckets.

This is preferable to synthesizing zero gradients or disabling
overlap_grad_reduce, because neither workaround changes the fact that the
current QSA indexer has no gradient path.

Tests

• Assert every QSA indexer parameter has requires_grad=False and grad is None.
• Assert all pretrained indexer parameters remain present in state_dict().
• Run the existing forward-only and QSA selection tests.