[Core][Modes] New buffer object to track underlying mode allocations …

[noelchalmers]

Description

This PR slightly changes how reference to backend memory allocations are counted so they can be more reliably destroyed, and prevent memory access violations via invalidated pointers.

The changes are entirely internal to OCCA and do not change the public API, or alter current behavior of OCCA applications.

Summary of issue:

Currently modeMemory_t objects do not track whether they were created by slicing another modeMemory_t object. Consequently, only the original modeMemory_t object can free the underlying device memory allocation, and will do so even if there still exist modeMemory_t objects referencing the device memory. This results in dangling references.

E.g.

{
  occa::memory o_a;
  {
    occa::memory o_A = device.malloc<float>(2*entries); //Create a new buffer
    o_a = o_A + entries; //  o_a is a slice into o_A
  } //o_A is destroyed here because it's reference counting is separate from o_a.
    // The underlying device memory is also destroyed

  // o_a is now invalid, but doesn't know the original device memory is free'd

  // Using o_a will result in an access violation
  myKernel(entries, o_a); //Fault

} //o_a destroyed here, but since it's not the origin, there's no double free

This PR addresses this behavior. The new buffer objects wrap device memory allocations and track their modeMemory_t references, created via slicing. The buffer and wrapped device allocation are destroyed once their reference count reaches zero.

[codecov]

Codecov Report

Merging #490 (618fa3a) into main (5028b14) will increase coverage by 0.00%.
The diff coverage is 97.32%.

❗ Current head 618fa3a differs from pull request most recent head f7d0371. Consider uploading reports for the commit f7d0371 to get more accurate results

@@           Coverage Diff            @@
##             main     #490    +/-   ##
========================================
  Coverage   76.24%   76.24%            
========================================
  Files         261      256     -5     
  Lines       19399    19275   -124     
========================================
- Hits        14790    14696    -94     
+ Misses       4609     4579    -30     

Impacted Files	Coverage Δ
src/core/device.cpp	92.43% <ø> (-0.04%)	⬇️
src/occa/internal/modes/serial/buffer.cpp	91.30% <91.30%> (ø)
src/core/memory.cpp	79.85% <91.66%> (-1.23%)	⬇️
src/core/kernelArg.cpp	81.73% <100.00%> (ø)
src/occa/internal/core/buffer.cpp	100.00% <100.00%> (ø)
src/occa/internal/core/device.cpp	75.00% <100.00%> (ø)
src/occa/internal/core/memory.cpp	100.00% <100.00%> (ø)
src/occa/internal/modes/serial/device.cpp	90.76% <100.00%> (-0.05%)	⬇️
src/occa/internal/modes/serial/memory.cpp	100.00% <100.00%> (ø)
include/occa/dtype/utils.hpp	66.66% <0.00%> (-33.34%)	⬇️
... and 25 more

[dmed256]

This PR is freaking great, glad we're making memory slices safer ❤️
I have a few comments but overall it looks great!