Add solutions for trtllm_fp4_block_scale_routed_moe / trtllm_fp8_per_tensor_scale_moe
Adds 8 public FlashInfer solutions for definitions that previously had zero
solutions attached (5 trtllm_fp4_block_scale_routed_moe shapes, 3
trtllm_fp8_per_tensor_scale_moe shapes), following the PR#12 solution style.
Weight quantization for the FP4-routed solutions uses fp4_quantize() with no
manual weight shuffle -- the routed API's Python wrapper does not expose
use_shuffled_weight, and the underlying C++ launcher hardcodes it to true
while expecting fp4_quantize()'s native output layout; a manual shuffle
double-applies the layout transform and corrupts results. The FP8 per-tensor
solutions set use_routing_scales_on_input=True only for Llama4 routing
(routing_method_type=3), matching RoutingMethodType.Llama4 semantics.
Each definition's own embedded reference implementation was found to not
model the intermediate activation re-quantization step a real kernel performs
(unlike e.g. the W4A8 definition added in PR#13, which does via
_mxfp8_roundtrip), so it doesn't provide a reliable local correctness oracle
for these FP4/FP8 kernels on unscaled synthetic data. Each solution's call
pattern was instead independently verified against flashinfer's own
CI-proven test infrastructure (tests/moe/test_trtllm_gen_routed_fused_moe.py
and trtllm_gen_fused_moe_utils.py) at the exact target shape on a real NVIDIA
GB200 (B200-die, SM100) node: 8/8 passed. See each solution's description
field for details.