perf(ep): optimize dispatch intranode kernel perf in cases with small tokens

[kawhil-amd]

This PR introduces a warp-group optimization to the IntraNode dispatch kernel (src/ops/dispatch_combine/intranode.hpp), enabling multiple warps to cooperatively process each token-expert pair for improved memory bandwidth utilization within small token condition.

Usage

To use the IntraNodeLL kernel, specify kernel_type=IntraNodeLL when creating the config:

import mori
import torch

config = mori.ops.EpDispatchCombineConfig(
    data_type=torch.bfloat16,
    rank=rank,
    world_size=8,
    hidden_dim=7168,
    max_num_inp_token_per_rank=64,
    num_experts_per_rank=32,
    num_experts_per_token=8,
    block_num=64,
    warp_num_per_block=4,
    use_external_inp_buf=True,
    kernel_type=mori.ops.EpDispatchCombineKernelType.IntraNodeLL,
)

op = mori.ops.EpDispatchCombineOp(config)

# Dispatch and combine work the same as IntraNode
dispatch_output, ... = op.dispatch(input_tensor, weights, indices, ...)
combine_output, ... = op.combine(combine_input, weights, ...)

Dispatch Latency (MI355X)

block_num	topk	max_token	Baseline(us)	Current(us)	Speedup
4	8	1	18.6	15.8	+15.08%
32	8	32	23.8	19.16	+19.4%
64	8	64	28.12	24.04	+14.6%

Test Plan

• test_dispatch_combine_ll passes for quant_type=none
• test_dispatch_combine_ll passes for quant_type=fp8_direct_cast
• test_dispatch_combine_ll passes for quant_type=fp8_blockwise

[isytwu]

Maybe you could record the data before and after optimization in this PR, as well as block/warp, for 128 and 4096 tokens, respectively?

[kawhil-amd]

Maybe you could record the data before and after optimization in this PR, as well as block/warp, for 128 and 4096 tokens, respectively?

Sure, will do it later.