Better CUDNN Dispatches

[avik-pal]

• Broadcasted Operations are often faster using cudnnOpTensor instead of the default .<op>
• We don't store boolean in state anymore. Using Val we can compile away most train vs test branching

[codecov]

Codecov Report

Merging #16 (a737ef5) into main (3e6a6d7) will increase coverage by 0.36%.
The diff coverage is 38.46%.

@@            Coverage Diff             @@
##             main      #16      +/-   ##
==========================================
+ Coverage   30.07%   30.44%   +0.36%     
==========================================
  Files          12       12              
  Lines         778      808      +30     
==========================================
+ Hits          234      246      +12     
- Misses        544      562      +18     

Impacted Files	Coverage Δ
src/Lux.jl	100.00% <ø> (ø)
src/transform.jl	79.31% <ø> (ø)
src/autodiff.jl	13.33% <8.33%> (+0.83%)	⬆️
src/nnlib.jl	38.83% <35.95%> (-9.61%)	⬇️
src/layers/basic.jl	41.37% <36.00%> (+0.23%)	⬆️
src/layers/normalize.jl	16.03% <36.66%> (-2.15%)	⬇️
src/layers/dropout.jl	37.03% <40.00%> (+4.77%)	⬆️
src/utils.jl	36.23% <61.53%> (+5.19%)	⬆️
src/layers/conv.jl	31.00% <85.71%> (ø)
src/core.jl	52.77% <100.00%> (ø)
... and 3 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 3e6a6d7...a737ef5. Read the comment docs.