Fix out-of-bounds check silently skipped for shape-only kernel params

[inducer]

check_bounds skipped the access-range-vs-shape subset test whenever an array's shape depended on a parameter that was not present in the instruction's domain space. A value arg that appears only in an argument shape (and not in the domain or in any instruction expression) is never a read dependency, so get_insn_domain never adds it to the domain, and the mapper's "shape_deps <= available_vars" guard bailed out without checking.

This let a kernel like

    { [i]: 0 <= i < 5 }   x[i, 0] = 5   x: shape (n, 2)

generate code even though the access is only in-bounds when n >= 5, and it did so without warning. ISL could not even express the needed premise because the parameter was absent from the space.

Fix in _AccessCheckMapper.map_subscript: when a shape dependency is missing from the domain space but is a read-only value arg of the kernel, add it as a free parameter and continue with the check. Other missing dependencies (e.g. temporaries) remain data-dependent and are still skipped. The result:

• without an assumption, the check now correctly raises LoopyIndexError;
• with the assumption n >= 5, it passes, as expected.

Also teach assume()'s string form to declare value-arg names (not just domain params) as assumption parameters, so lp.assume(knl, "n >= 5") works for shape-only params.

Add test_bounds_check_with_shape_only_param covering both the error and the assumption-pass cases.

Note: kernels where the shape param is a different domain param (e.g. domain 0<=i<m with shape (n,)) now fail unless an assumption such as m <= n is given -- the intended soundness improvement.

Assisted-by: Zed:Qwen3.8-27B-FP8