Support grid-constant by-value kernel parameters for TMA descriptors #1223

What problem does this solve?

CUDA recommends passing a tensor map to a kernel as a read-only, by-value
__grid_constant__ parameter. cuda-oxide currently exposes TMA operations over
*const TmaDescriptor, so callers that need a runtime descriptor must allocate
device memory for its 128 bytes, copy the encoded descriptor to that allocation,
and pass an ordinary device pointer. This adds storage and transfer work and
does not express the intended immutable kernel-parameter ABI.

Proposed API / behavior

Provide one Rust kernel-parameter declaration whose device ABI and generated
host launcher both derive from it. The exact spelling is open to design; for
example:

#[kernel]
pub unsafe fn copy(
    input_tma: GridConstant<TmaDescriptor>,
    // ...
) {
    cp_async_bulk_tensor_2d_g2s(
        dst,
        input_tma.as_ptr(),
        coord0,
        coord1,
        barrier,
    );
}

The implementation would:

• lower the aggregate kernel parameter to a pointer-typed LLVM parameter with
  byval(TmaDescriptor) and the required alignment;
• add that parameter index to the kernel's grid_constant entry in
  !nvvm.annotations;
• keep the device view read-only and addressable so it can be passed directly
  to TMA intrinsics;
• make #[cuda_module] marshal the descriptor bytes as one by-value kernel
  argument, rather than requiring a device allocation;
• derive both sides from the one parameter declaration rather than adding an
  independent launcher option.

Ordinary pointers to tensor maps in global memory should remain available for
the genuinely mutable/dynamically selected descriptor case.

Alternatives considered

• Allocate a 128-byte device slot, copy the descriptor H2D, and pass its
  pointer. This works today but adds storage/transfer work and a separate
  lifetime to retain.
• A constant-memory global is not per invocation and conflicts with concurrent
  kernels carrying different descriptors.
• Inline PTX cannot change an already-established kernel parameter ABI.

Additional context

• CUDA Programming Guide, TMA tensor maps:
  https://docs.nvidia.com/cuda/cuda-programming-guide/04-special-topics/async-copies.html
• CUDA __grid_constant__ parameter semantics:
  https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/cpp-language-extensions.html#grid-constant
• NVVM IR grid_constant property: the annotated parameter must be a pointer
  with byval, and the property value lists 1-based parameter indices:
  https://docs.nvidia.com/cuda/nvvm-ir-spec/#supported-properties

cuda-oxide already has aggregate-by-value launch machinery and emits other
kernel properties through nvvm.annotations; this request is for the missing
end-to-end parameter contract, including host marshalling.