feat: add Apple Intel CPU + MPS support

[amul3]

Adds backend auto-detection (cuda > mps > cpu) so the training script runs on Apple Intel Macs without a discrete GPU and on Apple Silicon, in addition to the original NVIDIA CUDA path.

Key changes:

• train.py: lazy/optional Flash-Attention 3 import (CUDA only); SDPA fallback for MPS/CPU; conditional torch.compile (CUDA only); conditional cuda.synchronize / cuda.manual_seed / max_memory_allocated; fp32 embeddings + rotary on non-CUDA backends to avoid slow/unsupported bf16 matmul; per-backend default DEPTH / DEVICE_BATCH_SIZE / TOTAL_BATCH_SIZE / WINDOW_PATTERN so the 5-minute budget actually completes on CPU.
• prepare.py: dataloader buffers allocate on detected device; pin_memory only on CUDA; evaluate_bpb pulls device from the model rather than hardcoding cuda.
• pyproject.toml: drop hard kernels dep + CUDA-only torch index; move kernels to an optional [cuda] extra; loosen torch to >=2.4 so CPU/MPS wheels resolve.
• README: document the new platforms, install matrix, and per-backend defaults.