[Feature] Add query-chunked TileLang DSA selector

[Denny991]

Motivation

The GLM-5.2 DSA Indexer selector originally processed all query tokens in a
single TileLang launch. For long sequences, this creates a large temporary
dense score buffer and increases peak GPU memory usage.

This PR introduces query chunking for the selector. Query tokens are processed
in independent chunks, while the fixed-ID Top-K output semantics remain
unchanged.

Changes

• Add optional query_chunk_size support to the DSA selector interface.
• Process TileLang selector queries chunk by chunk.
• Preserve packed-sequence causal ranges for every query chunk.
• Safely pad non-aligned query tails for TileLang kernel constraints.
• Keep only the final contiguous torch.int32 Top-K ID tensor.
• Use an 8K query chunk by default in the GLM-5.2 recipe.
• Keep the PyTorch selector unchanged.
• Keep the existing SparseMLA and cuDNN forward/backward kernels unchanged.
• Add unit tests for chunking, tail padding, short sequences, and one-shot
  versus chunked Top-K parity.

Usage

The GLM-5.2 recipe uses an 8K query chunk by default.

To select a custom chunk size:

export INDEXER_TOPK_QUERY_CHUNK_SIZE=2048

To disable query chunking:

export INDEXER_TOPK_QUERY_CHUNK_SIZE=0

Low-level selector APIs still use None as their default, so existing callers
are not changed implicitly.

Behavior

• Query length smaller than 8K is processed in a single chunk.
• Query length larger than 8K is split into multiple query chunks.
• The final output keeps the same shape and dtype:
  (seq_len, 1, topk) and torch.int32.
• Invalid causal positions remain represented by -1.
• The chunk size only limits temporary selector workspace; it does not change
  the selected Top-K IDs.

Memory Impact

The temporary dense selector score buffer changes from [S, S_k] to
[C, S_k], where C is the query chunk size.

For S = S_k = 16K:

Query mode	Chunk size	Dense score workspace
One-shot	None	approximately 1.00 GiB
Chunked	8192	approximately 512 MiB
Chunked	2048	approximately 128 MiB

These numbers describe the selector workspace only. Total model peak memory
also includes activations, communication buffers, and other temporary tensors.

Validation

• Targeted selector and attention tests passed on 8× NVIDIA H200.
• One-shot and chunked selectors produced identical Top-K IDs on packed inputs.
• Non-aligned query tails and short-query cases passed.
• Reduced-model GLM-5.2 E2E training completed with packed 16K input and a
  2K query chunk.
• Losses remained finite, with no OOM, CUDA, NCCL, or traceback errors.
• Python compilation and git diff --check passed.

Limitations

• Query chunking is implemented for the TileLang-backed DSA selector.
• The DSA loss implementation is not modified.
• Existing GLM-5.2 batch-size and parallelism constraints remain unchanged.