Changed `LQJacobianOperator` to store jacobian as 3 tensors and added support for `CUDA.jl`

[dlcole3]

• The LQJacobianOperator now stores the implicit Jacobian as 3 tensors rather than 1 matrix (this makes the block matrices contiguous in memory). I also updated the mul! functions and add_jtsj! to work with these tensors (mul! functions now take 2 for loops rather than 1).
• The tensors use a new function _dnlp_unsafe_wrap to access the data from the tensors (results in less allocations than calling view within LinearAlgebra.mul!).
• Added CUDA as a dependency. In the future, we may want to make the GPU functions a separate package/extension
• Added new mul! and add_jtsj! functions that work with CuArrays. These functions are defined with multiple dispatch, and the functions for calculating $Jx$ and $J^T x$ use gemm_strided_batched!. This could be optimized more in the future
• add_jtsj! with CuArrays is not close to optimal. Unfortunately, mul!(C, A, B, alpha, beta) does not work with CuArrays when C is accessed with a view, so I had to define a new storage matrix, H_sub_block, within LQJacobianOperator. I expect this will change in the future when I update this function with CUBLAS functions.
• Added tests for making sure that add_jtsj! and mul! using the GPU give the correct answer.

[codecov-commenter]

Codecov Report

Merging #39 (966bb47) into main (5ff3dee) will decrease coverage by 0.25%. The diff coverage is 97.36%.

@@            Coverage Diff             @@
##             main      #39      +/-   ##
==========================================
- Coverage   97.56%   97.30%   -0.26%     
==========================================
  Files           1        1              
  Lines        1066     1186     +120     
==========================================
+ Hits         1040     1154     +114     
- Misses         26       32       +6     

Impacted Files	Coverage Δ
src/DynamicNLPModels.jl	97.30% <97.36%> (-0.26%)	:arrow_down:

Help us with your feedback. Take ten seconds to tell us how you rate us.