fix(moe): auto-degrade FP8 block_n/k from 128 to 64 on alignment mismatch (#442)

[haowu1234]

Problem

Models with intermediate_size=1536 (MiniMax M2.5, Qwen3-235B) at TP=8 produce intermediate_size_per_partition=192, which is not divisible by block_n=128 (the default for per_1x128 block quantisation). Both CompressedTensorsFp8MoEMethod and Fp8MoEMethod raise ValueError, forcing users to fall back to TP=4 and waste 4 GPUs.

The existing error:

ValueError: The output_size of gate's and up's weight = 192 is not divisible by weight quantization block_n = 128.

Additionally, after fixing block_n, block_k=128 also fails at TP=8 (192 % 128 != 0), which was masked by the earlier block_n error.

Fix (Option B from #442)

When intermediate_size_per_partition is not divisible by 128 but is divisible by 64, auto-degrade block_n and/or block_k to 64 and emit a logger.warning. The ValueError is preserved when alignment is truly impossible (e.g. 190 % 64 != 0).

Changes

• CompressedTensorsFp8MoEMethod.create_weights: auto-degrade self.block_n and self.block_k before alignment checks.
• Fp8MoEMethod.__init__: initialise self.block_n/self.block_k instance attributes (were local variables in create_weights only).
• Fp8MoEMethod.create_weights: auto-degrade both block sizes using instance attributes.
• Fp8MoEMethod.get_fused_moe_quant_config: use [self.block_n, self.block_k] instead of hardcoded [128, 128] that would mismatch the degraded scale shapes.
• tests/test_fp8_moe_block_align.py: 13 unit tests covering degrade path, aligned path, impossible alignment, per_1x32 no-op, scale shape correctness, block_k degradation at TP>1, and full TP=8/1536 end-to-end scenarios.

Validation

• 13/13 tests passed on MI300X (gfx942, ROCm 7.1, Python 3.12, aiter 0.1.1.dev1)
• scale shapes verified: block_n=block_k=64 → w13_scale=(E, 6, 112), w2_scale=(E, 112, 3) for 192 intermediate at TP=8
• existing TP=4 / aligned paths unchanged

Refs: #442