Fix batched cov transform

[tbensonatl]

Fix a few issues with batched cov() transforms:

• Use transpose_matrix instead of transpose for the batched transposed deviations. Thus, dimensions [0, RANK-2) match the sizes of the batched dimensions from the non-transposed deviations.
• Use stream-oriented async allocators for temporaries and set the memory space to MATX_ASYNC_DEVICE_MEMORY.
• Make some changes in dimension-checking for batched matmul() transforms. Compute the number of batches using the dimensions of C and, if necessary, broadcast A or B into the batched dimensions.
• Add a cov() unit test case with batching and rectangular inputs.

One remaining issue is that we allocate 3-4 times the size of the input for temporary storage. This storage is part of a cached plan that currently is not evicted/freed, so if the dimensions of cov() inputs are changing over time, the cache will grow quickly.

[cliffburdick]

/blossom-ci

[cliffburdick]

@tbensonatl looks like an example was removed:

Warning, treated as error:
/home/jenkins/workspace/unit-tests/docs_input/api/manipulation/rearranging/transpose.rst:15:start-after pattern not found: example-begin transpose-test-1

[tbensonatl]

/blossom-ci

[tbensonatl]

/blossom-ci