Optional GPU training via CUDA backend (--features cuda)

[bddap-bot]

Optional GPU training (CUDA)

Adds a cuda Cargo feature that moves training/inference onto the GPU
(cubecl-cuda) on the RTX 2080. CPU (ndarray) stays the default — it's faster
for the current small model and needs no GPU or toolkit. Enable GPU with
cargo run --features cuda.

Why not wgpu

wgpu was the portable first choice and a dead end on burn 0.20: its wgpu module
is !Send (wgpu-core 26's PendingWrites is !Sync), but both
AutodiffModule::valid() and Optimizer::step require Send, so
Autodiff<Wgpu> won't compile. The CUDA backend is Send-clean and works as-is.

Design

• Cargo.toml: cuda = ["burn/cuda"]; ndarray stays on always, so cargo test and the default build need no GPU.
• session.rs: TrainBackend is cfg-selected (ndarray | cuda). InferBackend
  and Device derive from it via associated types, so the rest of the code is
  backend-agnostic — no cfg scattered through the logic.
• shell.nix: allowUnfree + cudaPackages.cudatoolkit + CUDA_PATH, and the
  host driver lib (/run/opengl-driver/lib, where libcuda lives) appended to
  LD_LIBRARY_PATH.

Measured (honest)

Headless training, both backends learn identically (reward 866→1010, crab stands,
every episode runs the full 500 steps):

backend	steady steps/s
CPU (ndarray)	~670
GPU (cuda)	~500–530

GPU is slower at this scale — single env, a 256-wide MLP, m=1 matmuls, so
kernel-launch overhead dominates and there's a one-time ~30 s kernel autotune.
The GPU win arrives with bigger models and batched/parallel envs, where the
matmuls become compute-bound. This lands the capability now so scaling up is a
flip of a feature flag, not a porting project.

Verification

• Default (CPU): cargo fmt --check, clippy --all-targets -D warnings,
  cargo test (5) — green.
• --features cuda: cargo check + clippy -D warnings — green; training
  verified live on the 2080 (cubecl initialises device 0, JIT-compiles kernels).
• Built/run in a sandbox; the GPU run used a locked-down bwrap that exposes the
  NVIDIA devices + driver while keeping $HOME/secrets hidden.

Not bundled here: per-episode pose metrics (height/uprightness) and a way to
render a trained crab — next up, to make iterating measurable.