Add ep_size to ParallelismConfig so FSDP2 can compose with expert parallelism

[YeonwooSung]

What does this PR do?

ParallelismConfig had dp_replicate, dp_shard, tp, cp, and sp, but no expert-parallel axis. A 2-way FSDP × 4-way EP layout on 8 GPUs could not be expressed, and every downstream stack had to invent its own mesh.

This adds ep_size (env PARALLELISM_CONFIG_EP_SIZE, CLI --parallelism_config_ep_size, accelerate config when FSDP2 parallelism is enabled). Default is 1.

• total_size and non_data_parallel_size include ep_size
• "ep" is in non_dp_dim_names when ep_size > 1 (same batch on every EP rank, like TP/CP)
• "ep" is not in fsdp_dim_names / dp_shard_cp_dim_names, so FSDP2 does not double-shard expert weights
• Mesh order: dp_replicate, dp_shard, cp, sp, ep, tp (only enabled dims appear)
• prepare_data_loader treats EP like TP/CP when mapping ranks onto data-parallel groups

Expert placement stays with the model. Accelerate only builds accelerator.torch_device_mesh["ep"].

cfg = ParallelismConfig(dp_shard_size=2, ep_size=4)
accelerator = Accelerator(
    fsdp_plugin=FullyShardedDataParallelPlugin(fsdp_version=2),
    parallelism_config=cfg,
)
# mesh["ep"] is the expert group; mesh["dp_shard"] is the FSDP group

Fixes #4200. Config/mesh half of the FSDP2 × EP work around #4178 / huggingface/transformers#48204 / huggingface/trl#6869.

Not in this PR: token-choice dispatch, FSDP2 wrap-heuristic changes, USP (CP+SP).

Tests

• Mesh shapes for EP-only, FSDP+EP, FSDP+EP+TP, HSDP+EP
• EP is not flattened into FSDP dims
• Env PARALLELISM_CONFIG_EP_SIZE
• ep_size=0 raises
• Default ep_size=1 does not change existing CP/TP meshes
• dp_replicate>1 + ep>1 is allowed
• DataLoader rank mapping keeps one batch per EP group

Who can review?

@SunMarc @qgallouedec