This PR enables configuring OpenXLA allocator settings on ROCm targets and ensures that ZML respects user-defined allocator overrides set via standard environment variables.

Previously, the rocm platform target configuration in ZML was a stub (struct {} = .{}, which wrote no PJRT client NamedValue options). On AMD ROCm targets (specifically unified memory APUs like AMD Strix Halo), the ROCm PJRT driver (libpjrt_rocm.so) falls back to its default behavior of preallocating 90% of available memory. This frequently triggers system-level Out-of-Memory (OOM) events when running concurrent compilation or serving workloads, even if the user configures lower limits in their environment.

This change:

• Updates writeNamedValues to check if standard XLA environment variables (XLA_CLIENT_PREALLOCATE and XLA_CLIENT_MEM_FRACTION, along with their XLA_PYTHON_* equivalents) are set.
• If they are set, ZML bypasses serialization for those options, allowing the underlying OpenXLA engine to parse and apply the environment variable values directly.
• Falls back to bfc allocator when async is requested on ROCm (since cuda_async is CUDA-only).
• Preserves configured code options when no environment overrides are present.
• Adds a new unit test CreateOptions.toNamedValues allocator config covering all environment variable bypass and fallback cases.

Platform Impact

• ROCm: Resolves driver default-override limitations and prevents unified memory OOM crashes under restricted UMA/GTT carve-outs.
• CUDA: Unifies the configuration behavior with ROCm and ensures environment overrides work consistently.

Verification

Compiled and verified the core ZML tests using:

./bazel.sh test //zml:test