Compiler optimizations

[pgbrodrick]

Leverage views and strategic function definition for fully type-specific compilation.

In practice, this reduces runtime by about 35% (EMIT style runs) with identical results. This PR uses more sophisticated views and TLS, but no increased performance was found, and there is notably more complexity, so keeping this implementation. The solver Levenberg-Marquardt solver referenced in the above is perhaps worth considering, however.

Note that the above PR also identifies a small algorithm bug in the BVLS solver, which I chose not to address here. Presently, we modify all bounds simultaneously, instead of one at a time. Technically, I believe this is incorrect, but in practice I found minuscule differences from the unmixing problem. For the sake of numerical consistency, I've allowed it to persist, though this could be reconsidered.

[vatsal-j]

@pgbrodrick The code itself looks good as far as I can tell (and you've already tested it). The doc building test failed because the new functions:

│     SpectralUnmixing._unmix_pixel_kernel :: Tuple{SpectralLibrary, Matrix{Float64}, Union{Nothing, Matrix{Float64}}, Any, Any, String, Int64, Vector{Int64}, String, String, Int64, String}
│     SpectralUnmixing.ensure_2d :: Tuple{AbstractVector{Float64}}
│     SpectralUnmixing.ensure_2d :: Tuple{AbstractMatrix{Float64}}

need to be added to docs/src/api.md

[pgbrodrick]

Merging now that the above request on documentation has been cleared.