[Software pipeline] Fix hardcoded index in access_ptr rewriting, add a GPU test with depth 4

[masahi]

Fix a hardcoded index in access_ptr rewriting, which assumes that the number of stages is 2.

Refactored MMA code in test_tir_schedule_tensorize_ldmatrix_mma.py, so that it can be used by other tests. The new test in test_tir_transform_inject_software_pipeline.py applies software pipelining annotations to the MMA-tensorized schedule with software_pipeline_stage = [0, 0, 3], which makes global to shared load pipelined with depth 4. Without async copy, this is not useful for performance. But it does demonstrate that a multi-stage pipeline with depth > 2 works on a semi-realistic GPU schedule.

The test uses large dynamic shared memory, which serves as a test case for #11478.

@vinx13 @junrushao1994 @csullivan