Add torch tensor cache to conv2d unit tests to speedup test execution (…

…#17708)

The function `torch.randn` takes a significant amount of time while
executing `tests/ttnn/unit_tests/operations/test_new_conv2d.py`. The
idea is to cache torch tensors with specific dimensions and reuse them
in other tests that require tensors of those dimensions. It turns out
that, out of approximately 3000 tensors that needed to be generated,
there were only around 300 unique dimensions. This approach reduces the
test execution time by 10%.