0.5.0

This will likely be one of, if not the last release of pre-1.0 Scalix (see
Roadmap.md for more information). Up to this point, I've been adding features
and fixing bugs as I've needed them for my own research. However, that project
is coming to a close so Scalix features will likely be paused until the paper is
submitted for publication.

With the project I've been working, a highly parallel fluid simulation method
for acoustics, the initial hope was to have a distributed implementation across
many GPUs, hence Scalix was born. Unfortunately, the reliance of Scalix on CUDAs
unified memory management ended up being a bottleneck for the project, so the
distributed implementation was never used, despite having automatic support via
Scalix's opaque distributed API, because one particular part of the method was
not conducive to CUDAs managed memory heuristics.

Using Scalix extensively for this project has been incredibly useful, but the
pre-1.0 APIs are clunky, requiring complicated workarounds for use cases that
weren't considered upon initial design. Rather than continuing pre-1.0
development, I will begin development of the 1.0 version, which will use SYCL as
a base. Using what I've learned from my usage so far, I'm much more confident in
the design of the 1.0 version, and I'm excited to get started. Some proof of
concepts have already been explored and the performance is looking promising.

Without further ado, I present tha changes for the final pre-1.0 release of
Scalix, except for possible bug fixes. Next major entry in this log will be a
beta release of 1.0!

Features

• Serialization support for sclx::array types using cereal
• STL-like methods:
  • sclx::algorithm::count_if
  • sclx::algorithm::min_element/max_element
  • sclx::algorithm::elementwise_reduce now supports concurrent execution'
  • sclx::iota
• Kernel info now provides information about the grid stride loop, useful for
  prefetching
• better sclx::array constructors, avoiding macro guards for host vs device
  implementations
• Better implicit casting of sclx::array non-const types to const types

API Changes

• Lots of changes of pass-by-reference to pass-by-value for sclx::array types.
  Not only is the cost of copying shared pointers minimal compared to the
  typical computational cost of various numerical algorithms, but passing by
  value improves thread safety.
• sclx::cexpr_memcpy is now sclx::constexpr_assign_array, better aligning
  with what the function actually does

Bug Fixes

• sclx::local_array now aligns allocations to its respective type's alignment,
  alleviating a rare bug that caused CUDA kernels to fail with a cryptic error
  message

Full Changelog: 0.4.1...0.5.0