[TP] Keep optimizer state aligned after parameter swaps

[sdjasj]

What does this PR do?

Keeps optimizer state keyed by the active parameters when tensor parallel preparation replaces model parameters with DTensors.

Optimizers such as Adagrad initialize per-parameter state in their constructor. TP preparation previously updated only param_groups, leaving the state dictionary keyed by the original parameters. PyTorch then raised a KeyError when serializing the inconsistent optimizer.

AcceleratedOptimizer._switch_parameters now moves any existing state to the replacement parameter, and TP preparation uses that state-aware path. The regression test verifies both the state keys and values before calling state_dict().

Fixes #3855

Validation

CUDA_VISIBLE_DEVICES='' python -m pytest -q tests/test_optimizer.py
# 2 passed, 1 skipped

# Four-GPU TP smoke test with a partially DTensor-backed model and Adagrad
accelerate launch --multi_gpu --num_processes 4 --main_process_port 0 /tmp/repro-accelerate-3855.py
# Adagrad optimizer state follows TP parameters

ruff check src/accelerate/accelerator.py src/accelerate/optimizer.py tests/test_optimizer.py
ruff format --check src/accelerate/accelerator.py src/accelerate/optimizer.py tests/test_optimizer.py