Run whole test suite on numba backend

[ricardoV94]

What is broken / not supported by Numba:

• float16: meta issue: fp16 support numba/numba#4402
• Deepcopy / copying protected outputs: Numba backend does not deepcopy shared outputs #50
• Sparse operations (progress in Implement Sparse Ops in Numba #1676)
• Nested typed lists as global constants (minor)

[ricardoV94]

Locally I ran all the tests in tests/tensor/rewriting/test_basic.py twice (to allow caching) before and after making Numba default.
Before: 6s
After: 1m34s

Numba caching is baaaaaaaaaaaaaaaaaaaaaaaaaaaaaaad

[ricardoV94]

Some fun looking kernel crashes, such as in tests/compile/function/test_types.py::TestFunction::test_copy_share_memory or tests/tensor/rewriting/test_elemwise.py::TestFusion::test_elemwise_fusion[case72]

Both seem to be related to Elemwise (CC @aseyboldt)

[aseyboldt]

Found a small reproducer for one of the aborts:

x = dvector("x") out = 1 / x**2 graph = FunctionGraph(outputs=[out]) func = numba_funcify(graph)

It works if I run it through the fast_run rewrites, but fails without them.
I'll try to figure out more tomorrow.

[ricardoV94]

Didn't we find an issue with power specifically recently?

[ricardoV94]

numba/numba#9554

[ricardoV94]

Okay, I added a patch for the power crash, disabling fastmath if the power is an integer for now

[ricardoV94]

After caching the test file now runs in 50s after caching vs 6s before the PR, so only 8x slower now :(

[ricardoV94]

tests/tensor/test_slinalg.py::TestSolveTriangular::test_solve_triangular_grad is failing with unit_diagonal=True with the numba backend. Can you have a look @jessegrabowski ?

[ricardoV94]

Also need to let the complex solves fallback to obj mode