TST: relative tolerance on tile-interpolation AHE tests should be much lower

[neutrinoceros]

Originally posted by @neutrinoceros in #375 (comment)

[neutrinoceros]

I note that the current test is comparing brute-force AHE to tile-interpolated AHE with the same tile size, but tile interpolation uses double that tile size as the effective-contextual-region , which, for small and noisy images, may have an important impact on results.
I should primarily test against scikit-image's own implementation of tile-interpolation for correctness.