[Perf][Linear-Attn] Add FLA-compatible Gated DeltaNet prefill op

[superAngGao]

Motivation

GatedDeltaNetFwdOp is useful as a chunkwise/training-style forward primitive, but using it directly as an inference prefill operator has two user-visible issues.

1. The serving prefill contract is not aligned with mainstream linear-attention libraries such as FLA. In practice, users commonly have q/k/v in BTHD layout ([B, T, H, D]) and g/beta in [B, T, H], and expect (o, final_state) where o is BTHD and final_state is [B, H, DK, DV]. Reusing the current forward path requires extra layout adaptation and exposes training/backward-oriented intermediates.

2. The current forward path is not optimized for prefill. On Qwen-style Gated DeltaNet inference shapes, prefill needs a dedicated kernel decomposition and dispatch policy. The goal should be an FLA-compatible prefill interface whose performance is competitive with, and ideally faster than, FLA.

Proposal

Add a dedicated GatedDeltaNetPrefillFwdOp with:

• FLA-compatible BTHD as the main path:
  • q/k: [B, T, H, DK]
  • v/o: [B, T, H, DV]
  • g/beta: [B, T, H]
  • final_state: [B, H, DK, DV]
• Backward-compatible TileOPs head-major layout support for existing users.
• A zero-state serving prefill contract: (q, k, v, g, beta) -> (o, final_state).
• Dedicated prefill kernels and dispatch rather than wrapping the chunkwise forward op.
• Benchmark coverage against FLA on representative Gated DeltaNet prefill shapes.

Acceptance criteria

• Public op API supports the FLA-compatible BTHD contract.
• Existing TileOPs-style layout users keep a supported path.
• Correctness tests cover BTHD vs head-major layout equivalence and final-state output.
• Benchmarks show the dedicated prefill path is faster than FLA on the targeted prefill workloads.