gpuResampSlc.cu: fix out of bounds reading bug in transformTile

[scottstanie]

(note: have not been able to recompile this on the GPU machine to fully test. just makes logical sense now)

I was getting the following error sporadically when running CUDA versions of geo2rdr + gpuResampleSlc

  Reading in image data for tile 38 of 39
  Interpolating  tile 38 of 39
  terminate called after throwing an instance of 'isce3::cuda::except::CudaError<cudaError>'
    what():  Error in file .../cxx/isce3/cuda/core/gpuLUT1d.cu, line 77, function isce3::cuda::core::gpuLUT1d< <template-parameter-1-1>
  >::~gpuLUT1d() [with T = double]: cudaError 700 (an illegal memory access was encountered)
  zsh: IOT instruction (core dumped)

I thought it was a "my image is too big problem", since shrinking the linesPerBlock worked (sometimes). Other times, shrinking it didnt work, but just changing it did.

The CPU version in cxx/isce3/image/ResampSlc.cpp:322-327 has the correct bounds check:

      if ((iRowResampled < chipHalf) ||
              (iRowResampled >= (inLength - chipHalf)))
          continue;
      if ((iColResampled < chipHalf) || (
                  iColResampled >= (inWidth - chipHalf)))
          continue;
``
However, the GPU version has only `>`.
The chip reading loop iterates iChipRow from 0 to chipSize - 1 (0 to 8), accessing row:
```c++
    iTileRow = iRowResamp + iChipRow - chipHalf

The maximum iTileRow is iRowResamp + 8 - 4 = iRowResamp + 4. With the current bounds check (iRowResamp + 4 > inReadableLength, it will skip), the kernel proceeds when iRowResamp = inReadableLength - 4. That gives:

    max iTileRow = (inReadableLength - 4) + 4 = inReadableLength  // out of bounds

The valid row indices are 0 to inReadableLength - 1, so this reads one row past the end of the tile buffer.