Fix cuTENSOR contractions not working for FP16 inputs

[thomasfaingnaert]

Contractions for FP16 inputs with Volta Tensor Cores (typeA = CUDA_R_16F, typeB = CUDA_R_16F, typeC = CUDA_R_16F, typeCompute = CUTENSOR_COMPUTE_32F) didn't work before, due to a missing case in scalar_types in lib/cutensor/wrappers.jl:

ERROR: KeyError: key (Float16, Float32) not found
Stacktrace:
 [1] getindex(h::Dict{Tuple{DataType, DataType}, DataType}, key::Tuple{DataType, DataType})                                                                  
   @ Base ./dict.jl:482
 [2] plan_contraction(A::Union{CuArray, Array}, Ainds::Vector{Char}, opA::CUDA.CUTENSOR.cutensorOperator_t, B::Union{CuArray, Array}, Binds::Vector{Char}, opB::CUDA.CUTENSOR.cutensorOperator_t, C::Union{CuArray, Array}, Cinds::Vector{Char}, opC::CUDA.CUTENSOR.cutensorOperator_t, opOut::CUDA.CUTENSOR.cutensorOperator_t; pref::CUDA.CUTENSOR.cutensorWorksizePreference_t, algo::CUDA.CUTENSOR.cutensorAlgo_t, compute_type::Type)                                               
   @ CUDA.CUTENSOR ~/.julia/packages/CUDA/YpW0k/lib/cutensor/wrappers.jl:310

This PR fixes that, and also updates the testsuite accordingly.

[maleadt]

Let's actually have this run the CUTENSOR test suite on CUDA 11.5 by rebasing after #1239 is merged.

[codecov]

Codecov Report

Merging #1238 (e3cbbba) into master (8caad95) will increase coverage by 0.01%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #1238      +/-   ##
==========================================
+ Coverage   80.17%   80.19%   +0.01%     
==========================================
  Files         119      119              
  Lines        8390     8390              
==========================================
+ Hits         6727     6728       +1     
+ Misses       1663     1662       -1     

Impacted Files	Coverage Δ
lib/cutensor/wrappers.jl	95.47% <ø> (ø)
src/accumulate.jl	100.00% <ø> (+2.94%)	⬆️

---

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 c903144...e3cbbba. Read the comment docs.