Tuple literals passed to generic functions fail during module hashing #2167

Bug Description

Passing a tuple literal directly to a generic @wp.func fails while hashing the kernel module. Assigning the same tuple to a local variable before the call works.

import warp as wp


@wp.func
def product(values: tuple[int, ...]):
    return values[0] * values[1]


@wp.kernel(enable_backward=False, module="unique")
def evaluate(out: wp.array[int]):
    out[0] = product((2, 3))


out = wp.zeros(1, dtype=int, device="cpu")
wp.launch(evaluate, dim=1, inputs=[out], device="cpu")
print(out.numpy())

Expected: [6].

Actual:

RuntimeError: Failed to determine type code for argument 'values' of function product:
Unrecognized type '(<class 'int'>, <class 'int'>)'

The traceback goes through Adjoint.get_references(), iter_call_callable_arg_targets(), Function.get_overload(), and get_type_code(). The literal's inferred argument type is a raw tuple of types, which the signature encoder does not recognize.

The equivalent kernel body succeeds:

values = (2, 3)
out[0] = product(values)

Generic calls should accept either spelling consistently, including during dependency discovery and module hashing.

Related: #2099, #2100, #2101, #2166. The reproducer uses flat integer tuples, indexed access, and disables backward generation, so it does not depend on nested tuples, argument unpacking, or autodiff. This also affects the natural grouped-input spelling proposed for #1813.

System Information

• Warp 1.20.0.dev0, main commit 0c1ed78f7.
• Python 3.12.3, Linux x86_64, CPU device.
• Reproduced before applying tuple fixes; failure occurs before native kernel execution.