Use CeedMallocArray for void pointers in CUDA/HIP QFunctionContext

Use backend functions for SyncArray in CUDA and HIP

gpu - remove 'quoted' operator assembly kernels

jit - add interface for adding additional jit source dirs

move include/ceed-jit-source to include/ceed/jit-source

jit - use relpath from include/ceed-jit-source for jit source files

tidy - small tidy fixes

Merge pull request #920 from CEED/jeremy/conversion

Explicit casting of vector sizes

gpu - fix error handling in size conversion

Merge pull request #925 from CEED/gpu-assemble

Add some matrix assembly support to GPU backends

Add fallback kernel for larger element sizes in GPU assembly

Add LinearAssemble CUDA reference implementation for low-order elements

Add LinearAssemble HIP reference implementation for low-order elements

MAGMA: simplify atomic add usage and reduce MAGMA header file usage

minor - explicit casting of vector sizes

minor - update copyright headers

opt - reduce memory usage in CeedOperatorLinearAssembleQFunctionCore_Opt

rstr - use CeedSize for l_size

vec - use CeedSize for vector lengths

Merge pull request #893 from CEED/natalie/more-hip-launch-bounds

HIP: add atomics flag and more kernel launch bounds for performance improvements

Change to more specific name for hip-gen block size function

qf - extend ctx read/write feature to qf

ctx - add read-only access for QFContext

backends/magma: fix pinned vs unpinned memory free bug

backends cuda-shared: fix launch bounds to avoid invalid z dimension

The typical max z dimension size of a thread block is 64 and we were
computing larger values (like 85) in some cases.