fix: pre-initialize cast kernels at buffer construction

cuModuleLoadData hangs on H100 CUDA 13 when called after CUDA graph
capture — the context state appears incompatible with module loading.
Pre-initialize the OnceLock during GpuReplayBuffer::new() when the
context is clean and no graphs have been captured.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/ml-dqn/src/gpu_replay_buffer.rs

@@ -256,6 +256,12 @@ impl GpuReplayBuffer {
         let update_batch_max = a32f(stream, 1, "update_batch_max")?;
         let update_max_merge = a32f(stream, 1, "update_max_merge")?;
 
+        // Pre-initialize cast kernels while CUDA context is clean.
+        // OnceLock + cuModuleLoadData requires a current CUDA context.
+        // If deferred to first PER update (after graph capture), the context
+        // may be in a state that blocks cuModuleLoadData on H100 CUDA 13.
+        get_cast_kernels(stream)?;
+
         Ok(Self {
             config, stream: Arc::clone(stream), kernels: k,
             states: s, next_states: ns, actions: a, rewards: r, dones: d, priorities: p,