Precompute A^-1/n and A^-n in DislocationCreep and DiffusionCreep

[albert-de-montserrat]

Computing A^-1/n or A^-n is really expensive, particularly during the pseudo-transient iterations where this is redundantly done thousands of times. In this PR we just precompute and store those values in the correspondent structs.

It results in a ~30% speedup in JustRelax.jl (example for a 64x64 model)

• Main branch:

julia> @b compute_viscosity!($(stokes, phase_ratios, args0, rheology, viscosity_cutoff)...)
    332.667 μs (21 allocs: 8.328 KiB)

• This branch

julia> @b compute_viscosity!($(stokes, phase_ratios, args0, rheology, viscosity_cutoff)...)
    255.375 μs (21 allocs: 8.328 KiB)

TODO: units make it type unstable in some Julia versions

[codecov]

Codecov Report

Attention: Patch coverage is 90.90909% with 2 lines in your changes missing coverage. Please review.

Files	Patch %	Lines
src/CreepLaw/DislocationCreep.jl	84.61%	2 Missing ⚠️

Additional details and impacted files

📢 Thoughts on this report? Let us know!