Replace Qwen3.5 FLA paths with Attention Gym

[drisspg]

Human Note

Agent note

This is a draft follow-up to pytorch#4363 and depends on the paged chunk GDN addition
being prepared on top of meta-pytorch/attention-gym#425.

Use Attention Gym for Qwen3.5 GDN training, batch-invariant backward recomputation, recurrent
execution, decode, and paged chunk prefill. The paged prefill path advances vLLM SSM cache slots
in place, eliminating the previous gather/chunk/scatter copies and preserving the shared
[slots, H, V, K] layout across prefill and decode.

Remove the now-unused FLA backend selection and direct dependency, normalize Q/K explicitly at
the Attention Gym boundary, and update debug configurations to the fused backend K = V = 128
contract while retaining two key heads for TP=2 coverage.

The remaining serving-side copy is the much smaller convolution history [N, W - 1, C] during
varlen prefill. Attention Gym currently has paged single-token convolution decode, but not a paged
varlen convolution prefill API.

Performance

GB200, fixed-pointer warm-cache microbenchmark, T=4096, N=32, HK=8, H=16, K=V=128,
BF16 QKV, FP32 state:

Path	Median GPU time
Caller gather + chunk + scatter	219.38 us
Direct paged chunk	185.95 us

Direct paging reduced this isolated subsystem time by 15.2%. This is not an end-to-end RL result.

Test Plan

PYTHONPATH=/home/drisspg/meta/attention-gym-pr425:/home/drisspg/meta/torchtitan pytest -q tests/unit_tests/gpu/test_qwen3_5_deltanet.py tests/unit_tests/test_qwen3_5_mrope_positions.py -x
PYTHONPATH=/home/drisspg/meta/attention-gym-pr425:/home/drisspg/meta/torchtitan pytest -q tests/unit_tests/cpu/test_state_dict_keys.py tests/unit_tests/cpu/test_integration_test_definitions.py tests/unit_tests/cpu/test_train_spec.py -x

The focused suites passed locally on GB200. Full RL/vLLM validation was not run because the local
environments do not contain vLLM. All configured formatting and lint hooks passed; repository-wide
Pyrefly remains blocked by unrelated missing optional dependencies and existing baseline errors.

[drisspg]

Closing this fork-only PR; the change is being resubmitted as a draft against pytorch/torchtitan on top of pytorch#4363.