codegen: support grid-constant kernel parameters - #1225

The generic-kernel path needs to be part of the design before this lands:

• For a generic #[kernel], the macro splits your fn into a generated #[inline(always)] helper that keeps the body and a concrete entry wrapper per instantiation that calls it. The marker lands in both. When rustc's MIR inliner folds the generated helper into the entry, detect_grid_constant_params hard-errors on the duplicate.
• Repro:

#[kernel] 
pub fn k<T: Copy>(#[grid_constant] d: &Desc, out: *mut u32, _tag: T)  { 
   unsafe { *out = d.values[3] } 
 } 

fails with kernel contains duplicate grid-constant marker for source parameter 0. Bigger bodies only compile because they exceed the inliner's cost threshold.

• More importanly, the helper's marker leaks: a plain kernel

other(d: &Desc, out: *mut u32) { 
  k::<u8>(d, out, 0) 
} 

with no attribute comes out with .param .align 64 .b8 other_param_0[128] and a grid_constant annotation. Host passes an 8-byte pointer, device expects 128 bytes by value.

• unchecked_indexing already handles this by stripping its marker from the re-emitted helper, see strip_unchecked_indexing_config_marker. grid_constant needs the same in both generic paths. A duplicate-tolerant detector alone fixes the error but not the leak.

And we need tests in the PR exercise a generic kernel with #[grid_constant].

To land this: the strip in both generic paths, a generic #[kernel] + #[grid_constant] test with a body small enough to be inlined, and a negative test that a non-opted kernel calling the helper still gets a plain pointer param.

Addressed the generic-helper ABI concern in f2b6b15.

The macro now captures the grid-constant marker for the generated entry and strips it from the re-emitted callable helper in both generic expansion paths. The contract example now covers both sides:

• an instantiated generic grid-constant entry retains the 128-byte by-value PTX parameter and launches successfully on sm_120;
• an ordinary kernel calling the same generic helper retains a .param .u64 pointer ABI and does not acquire the 128-byte parameter.

Measured validation:

• pixi run cargo test -p cuda-macros
• strict clippy for cuda-macros and cuda_module_contract
• pixi run cargo oxide run cuda_module_contract --arch sm_120 -- --verify-ptx
• pixi run cargo oxide run cuda_module_contract --arch sm_120
• the full just check code/test portion passed; its cargo-deny guard initially hit the local NFS advisory-lock limitation, then all deny/license guards passed with the advisory DB placed on local /tmp storage.