[AutoDeploy]: Improve MoE performance for TP>1

[nzmora-nvidia]

🚀 The feature, motivation and pitch

Autotuning was already enabled here: #8120

For tp=1, perf gap is ~8% on ToT.
For multiple GPUs we're still seeing a ~50% gap compared to manual trt-llm (previously was 300%-400%),

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.

[galagam]

Found two main issues:

1. non-optimial sharding config. set:

    use_sharding_from_factory: true
    support_partial_config: true

2. Use tp-sharding instead of ep-sharding in moe.
   weights shapes before:

w1_w3_stacked = (num_experts // world_size, 2*intermediate_size, hidden_size) 
w2_stacked = (num_experts // world_size, hidden_size, intermediate_size)

weights shapes after (same as TRT-LLM manual):

w1_w3_stacked = (num_experts, 2*intermediate_size//world_size, hidden_size)
w2_stacked = (num_experts, hidden_size, intermediate_size//world_size)

With both changes, AD is much closer to PT:
PT tp=2 62
PT tp=4 101
PT tp=8 (need to check)
AD tp=2 53
AD tp=4 98
AD tp=8 126

[galagam]

Perf improvement for Mixtral-8x7B-Instruct-v0.1, requests/sec.
Measured on cw-dfw (H100-80GB) with branch https://github.com/galagam/TensorRT-LLM/tree/user/galagam/moe-perf-sharding-fix

tp\version	main baseline	just cfg change	tp sharding
tp=2	28	49	53
tp=4	48	64	99
tp=8	Hanging	Hanging	120

Need to measure perf impact on other MoE architectures, and possibly control the sharding method via config / heuristic.