Fix command line parsing in the cuda.bindings example helpers

[LeSingh1]

Both helpers in cuda/bindings/_example_helpers/helper_string.py unpack enumerate()
backwards:

def check_cmd_line_flag(string_ref):
    return any(string_ref == i and k < len(sys.argv) - 1 for i, k in enumerate(sys.argv))


def get_cmd_line_argument_int(string_ref):
    for i, k in enumerate(sys.argv):
        if string_ref == i and k < len(sys.argv) - 1:
            return sys.argv[k + 1]
    return 0

enumerate() yields (index, value), so i is an int and k is a str — but the
body uses i as the argument text and k as the index. string_ref == i compares str
to int and is therefore always False.

$ python -c "from cuda.bindings._example_helpers import check_cmd_line_flag, get_cmd_line_argument_int; \
             print(check_cmd_line_flag('device='), get_cmd_line_argument_int('device='))" device= 3
False 0

So check_cmd_line_flag() always returns False and get_cmd_line_argument_int() always
returns 0. Every command line option in the examples is silently ignored — device=,
wA=, hA=, wB=, hB=, kernel=, help, ?, use_generic_memory — reaching
helper_cuda.find_cuda_device() / find_cuda_device_drv(),
3_CUDA_Features/global_to_shmem_async_copy.py, 0_Introduction/simple_zero_copy.py
and 2_Concepts_and_Techniques/stream_ordered_allocation.py.

Note the dead branch would not have worked either: k < len(sys.argv) - 1 is str < int
(TypeError) and sys.argv[k + 1] indexes with a str. Nothing here has ever run.

Three judgement calls, since "the current behaviour" is not a guide

Because both functions are constant today, any fix defines new behaviour. I kept it as
close to the evident intent — and to the C samples' helper_string.h — as I could:

1. check_cmd_line_flag() no longer requires a following argument. That condition
   belongs to the value lookup. Keeping it would leave help and ? broken whenever they
   are the last argument, which is the normal way to pass them.
2. Both helpers skip sys.argv[0], matching checkCmdLineFlag() in the C samples,
   which scans from argv[1].
3. get_cmd_line_argument_int() returns an int, as its name says and as its callers
   require: helper_cuda.find_cuda_device() passes the result straight to
   cudaSetDevice() and find_cuda_device_drv() to cuDeviceGet(). Returning
   sys.argv[idx + 1] unchanged would hand those APIs a str — i.e. fixing only the
   unpacking would turn a silent no-op into a TypeError. An int is also the only value
   this function has ever actually returned, since the literal 0 fallback was the sole
   reachable path.

Happy to change any of these if you'd prefer different semantics.

Tests

New cuda_bindings/tests/test_example_helpers.py — pure sys.argv monkeypatching, no GPU
or toolkit needed. Five of its assertions fail against upstream/main (flag present,
trailing boolean flag, flag in the middle of a longer argv, and both value lookups); the
program-name and not-found cases pass either way and are there to pin the new edges down.

tests/test_examples.py runs each example in a subprocess as [sys.executable, example],
so sys.argv[1:] is empty there and no flag can accidentally match under pytest.

Verified against upstream/main and with the change; ruff check and
ruff format --check are clean.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.