NVSHMEM4Py Integration

[Willy-Chan]

This PR will replace existing custom python bindings with NVSHMEM4Py - official Python language binding for NVSHMEM.

Scope of changes include:

• Migrate from static linking of host-side initialization (libnvshmem.a) to dynamic linking (libnvshmem_host.so)
• Remove existing bindings and replaced them with their respective NVSHMEM4Py core API
• Added helper function for NVSHMEM initialization

[Willy-Chan]

Confirmed functionality on the following system configurations:

1. DGX-H100 with Infiniband
2. DGX-H100 with AWS-EFA
3. B200 with MNNVL

These are benchmark perf difference results on dispatch and combine using MoEConfig(128, 8, 7186, 128):

System	Experts	E/Tok	Hidden Dim	Tokens	Dispatch Latency Perf Diff (lower is better)	Combine Latency Perf Diff (lower is better)
H100	128	8	7168	128	2.15%	2.96%
H100 (IB, 2 Node)	128	8	7168	128	0.84%	-2.94%
H100 (IB, 4 Node)	128	8	7168	128	0.59%	1.39%
H100 (IB, 8 Node)	128	8	7168	128	1.74%	-1.11%
H100 (EFA, 2 Node)	128	8	7168	128	-2.24%	1.27%
H100 (EFA, 4 Node)	128	8	7168	128	2.27%	-2.11%
H100 (EFA, 8 Node)	128	8	7168	128	-1.41%	-3.67%
B200	128	8	7168	128	1.08%	0.60%
B200 (NVLink, 2 Node)	128	8	7168	128	-2.23%	0.81%
B200 (NVLink, 4 Node)	128	8	7168	128	1.94%	-3.09%
B200 (NVLink, 8 Node)	128	8	7168	128	-5.60%	-2.90%

Performance percentages are calculated by measuring kernel latency using the provided benchmark, and negative percentage indicates that NVSHMEM4Py is faster.