Add SM70/SM75 forward GDN backend

[weicj]

Summary

This PR adds an explicit SM70/SM75 forward inference backend entry point for Qwen-style Gated DeltaNet.

The Hopper/SM90 TileLang path remains unchanged and stays the default. The new backend targets the Volta/Turing legacy family through a standard CUDA warp-shuffle implementation. Runtime validation of the standalone kernel was performed on RTX 2080 Ti / SM75; SM70 has compile coverage and still needs V100-class runtime validation.

Changes

• Add flash_qla.ops.gated_delta_rule.legacy.chunk_gated_delta_rule_fwd_legacy.
• Add a lazy-built CUDA extension for a forward-only legacy Gated DeltaNet backend.
• Keep the upstream Hopper/SM90 high-level API unchanged.
• Keep the legacy backend explicit instead of silently replacing the existing dispatch path.
• Add correctness coverage against a torch reference recurrence for the supported legacy path.
• Document the backend scope, limitations, opt-in behavior, and benchmark caveats.

Scope

Supported:

• forward inference only
• SM70/SM75-class CUDA devices as the intended legacy target family
• scalar-gate Gated DeltaNet
• Qwen-style grouped-query head mapping
• primary optimized shape: D=128
• explicit legacy API entry point

Not supported:

• backward kernels or training
• automatic dispatch from the upstream high-level API
• runtime performance claims for SM70 before V100-class validation
• SM80/SM86/SM89 support claims
• generic support for all pre-Hopper GPUs
• automatic default dispatch for non-Hopper devices

Evidence

Runtime validation of the standalone kernel was performed on RTX 2080 Ti / SM75. The same kernel code compiles for SM70, but SM70 runtime and performance validation is not claimed yet.

Standalone kernel timing for a Qwen-like shape:

• B=1, T=512, Hq=16, Hv=32, D=128
• control recurrent path: about 1.126 ms
• optimized legacy path on SM75: about 0.520-0.533 ms
• GDN-stage speedup: about 2.1x

GGUF runtime profiling on SM75:

• default fused GDN: 406.656 ms
• legacy fast path: 195.105 ms
• GDN-stage speedup: about 2.08x

Whole-request impact under the same server parameters:

• prefill: +7.17%
• decode: +0.61%
• wall time: -3.49%

The intended claim is limited to an explicit legacy forward path with measured GDN-stage improvement on SM75 and compile coverage for SM70. This PR does not claim a new end-to-end FlashQLA headline result.

Validation

• CUDA correctness tests added for supported legacy shapes.
• Python syntax check for the legacy wrapper and tests.
• Historical standalone RTX 2080 Ti / SM75 smoke benchmark for the source kernel.
• SM70 compile check.