Rename fused_marlin_moe kernels by precision scheme and split benchmarks

[zeroherolin]

PR Category

Operator | OP Test | Benchmark

Type of Change

Refactor

Description

Standardize fused_marlin_moe naming to fused_marlin_moe_<weight-act scheme>_<format> and split the benchmark by precision. Pure rename — no logic change, no new comments.

Kernels / entries / tune keys (src/flag_gems/fused/fused_marlin_moe.py + runtime registrations):

Before	After
fused_moe_w4a16_gptq	fused_marlin_moe_w4a16_int4
fused_moe_mxfp4	fused_marlin_moe_w4a16_mxfp4
_w4a16_* / _mxfp4_* kernels & invokes	_w4a16_int4_* / _w4a16_mxfp4_*
w4a16_pack / mxfp4_pack	w4a16_int4_pack / w4a16_mxfp4_pack
tune keys fused_marlin_moe_mxfp4[_gemm_silu]	fused_marlin_moe_w4a16_mxfp4[_gemm_silu]

INT8 (w8a16_int8) reuses the generic wna16 fallback and has no dedicated kernel; w8a16_fp8 is reserved for future work.

Tests (tests/test_fused_marlin_moe.py): test functions renamed to test_fused_marlin_moe_{w4a16_int4, w8a16_int8, w4a16_mxfp4}.

Benchmark: benchmark/test_fused_marlin_moe.py split into three self-contained per-precision files (test_fused_marlin_moe_{w4a16_int4, w8a16_int8, w4a16_mxfp4}.py), following the test_fused_moe_<precision>.py convention; pytest marker unified to fused_marlin_moe.

Issue

N/A

Progress

• Change is properly reviewed (1 reviewer required, 2 recommended).
• Change is responded to an issue.
• Change is fully covered by a UT.

Performance

No perf impact expected (identifier-only rename). Verified on H20 (sm90): 119 unit tests collected, per-precision subset + guard tests all pass; renamed tune keys load correctly via libtuner.