Performance improvements + remove numba as dependency

[jorgensd]

• Use vectorized numpy functions to unroll dofmap
• Replace find_pos (a numba convenience function) with vectorized numpy operations
• Remove numba as a dependency
• All crucial loops have been replaced by numpy vectorized operations

[jorgensd]

@francesco-ballarin this PR should give an insane speedup to any checkpointing. The tests goes from 3 minutes to 36 seconds on CI (with a few new tests introduced). Running a problem locally (with a 30x30x30 unit cube), timing components of read_function

This branch (serial):

Write mesh and function 8.99e-01
Read mesh 9.13e-01
Read function 2.05e+00

Main (serial):

Write mesh and function 2.38e+01
Read mesh 9.29e-01
Read function 8.53e+01

This branch (2 processes)

Write mesh and function 7.95e-01
Read mesh 6.51e-01
Read function 1.55e+00

Main (2 processes)

Write mesh and function 1.34e+01
Read mesh 7.07e-01
Read function 4.84e+01