fix_random_seed: seed only the CPU torch RNG, not every CUDA device

[HenryVarro666]

Fixes #1564.

Implements what was agreed in the issue: fix_random_seed() no longer touches CUDA RNG state and seeds only the CPU torch generator (torch.default_generator.manual_seed), plus random, numpy and lhotse's uuid4() as before. (Following up on the thread since the PR mentioned there didn't land yet — happy to adjust if you had a different shape in mind.)

Why: torch.random.manual_seed() also seeds the RNG of every visible CUDA device in the process. fix_random_seed() runs inside DataLoader workers (worker_init_fn in lhotse/dataset/dataloading.py) and at arbitrary points in user scripts, where silently rewriting GPU RNG state interferes with the training loop's own seeding policy. CUDA seeding is now explicitly the training script's responsibility, and the docstring says so.

Evidence (2x H100 NVL, torch 2.8.0):

import torch, lhotse
torch.cuda.manual_seed_all(1234)          # training script's own CUDA seeding
before = torch.cuda.get_rng_state_all()
lhotse.fix_random_seed(42)                # e.g. from a worker_init_fn
after = torch.cuda.get_rng_state_all()
# master:      CUDA RNG state changed on GPUs [0, 1]
# this branch: CUDA RNG state unchanged

Added two tests: CPU determinism is preserved (fix_random_seed → identical torch.randn draws), and a CUDA-marked test asserting torch.cuda.get_rng_state_all() is unchanged by fix_random_seed (passes on the 2-GPU machine, skipped without CUDA).