Keep ptxas resource counts scoped to the first kernel report - #12

When a cubin contains multiple kernels, parse_ptxas can combine the first kernel's register count with static shared memory or spills from a later kernel. For example, a real CUDA 13.0 SM103a compile reports 10 registers and no static SMEM for plain, then 128 bytes SMEM for shared; parsing the combined log incorrectly assigns 128 bytes to plain.

Scope matching to the first entry-function report, as the API already documents. Logs without compile headers use the first function-properties block; abbreviated single-report logs remain supported. Preserve absent fields rather than filling them from another function.

Validation:

• Focused tests with pytest -q -n 2 -p no:cacheprovider test_dump_kernel.py against isolated source-package overlays: base 2 failed, 3 passed, patched 5 passed. Both new tests fail without this change.
• Real nvcc -cubin -arch=sm_103a -Xptxas=-v and cuobjdump --dump-sass on a B300 environment, CUDA 13.0.88. Compiled the plain/shared pair in both declaration orders: the original parser mixes fields when the first report lacks SMEM; the patched parser equals the isolated first report in both orders.
• This is a resource-report correctness fix. It does not change generated kernels or claim a kernel speedup. The full NumSim/GPU suite was not run for this text-parser change.

The native reproducer uses these kernels (put shared before plain to obtain the plain report first with this toolkit):

extern "C" __global__ void shared(float* y, const float* x) {
    __shared__ float s[32];
    s[threadIdx.x] = x[threadIdx.x];
    __syncthreads();
    y[threadIdx.x] = s[(threadIdx.x + 1) % 32] * 3.0f;
}
extern "C" __global__ void plain(float* y, const float* x) {
    y[threadIdx.x] = x[threadIdx.x] * 2.0f;
}