Codegen slow due to excessive islpy calls

[View the loopy kernel on Gist.](https://gist.github.com/kaushikcfd/22485d07a101fb3dec6a2683eeee764d)

Profile log, as reported by `cProfile`, links 90% of the codegen time to islpy routines:

```
3024863 function calls (2382326 primitive calls) in 605.951 seconds

Ordered by: internal time

ncalls tottime percall cumtime percall filename:lineno(function)
790411/302110 477.026 0.001 477.870 0.002 islpy/islpy/__init__.py:861(wrapper)
265 83.808 0.316 83.884 0.317 islpy/islpy/__init__.py:1014(obj_eliminate_except)
18792/17160 11.141 0.001 11.482 0.001 loopy/loopy/isl_helpers.py:527(_find_aff_dims)
68 8.268 0.122 19.432 0.286 loopy/loopy/symbolic.py:1884(get_access_range)
74 6.624 0.090 233.594 3.157 loopy/loopy/isl_helpers.py:614(dim_max_with_elimination)
74 6.144 0.083 233.079 3.150 loopy/loopy/isl_helpers.py:609(dim_min_with_elimination)
4608 2.248 0.000 8.858 0.002 islpy/islpy/__init__.py:1100(_align_dim_type)
1070 1.455 0.001 2.364 0.002 loopy/loopy/symbolic.py:1590(__init__)
973/897 0.917 0.001 1.262 0.001 loopy/loopy/symbolic.py:1525(aff_to_expr)
35845 0.646 0.000 0.646 0.000 {built-in method islpy._isl.from_aff}
49830 0.596 0.000 1.861 0.000 islpy/islpy/__init__.py:1096(_set_dim_id)
```