[NDTensors] Optimize permutedims

[mtfishman]

This fixes a few performance issues in permutedims.

Inspired by this discourse question.

@kmp5VT

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (a00ef70) 84.05% compared to head (938a168) 53.94%.

❗ Current head 938a168 differs from pull request most recent head a9c355c. Consider uploading reports for the commit a9c355c to get more accurate results

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #1288       +/-   ##
===========================================
- Coverage   84.05%   53.94%   -30.12%     
===========================================
  Files         100       99        -1     
  Lines        8542     8489       -53     
===========================================
- Hits         7180     4579     -2601     
- Misses       1362     3910     +2548     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[kmp5VT]

It looks like there is an error with Strided and Base.ReshpaedArray. Heres the example

y = reshape(randn(elt, 8)', 2, 4)
x = Base.ReshapedArray(randn(elt, 8, 8)'[1:8], (2, 4), ())
y .= y .* x
# works 
@strided x .= x .* x

# stack overflow
@strided y .= y .* x
@strided x .= x .* y
@strided y .= y .* y

so I suggest we make a fix like

function permutedims(Edest::Exposed{<:Array, <:Base.ReshpaedArray}, Esrc::Exposed{<:Array}, perm, f) 
  adest = unexpose(Edest)
  asrc = unexpose(Esrc)
  adest .= f.(adest, permutedims(asrc, perm))
  return adest
end

and I can make an issue in Strided with my example above

[mtfishman]

That issue should be fixed by Jutho/StridedViews.jl#5.