GCInfo: Save bits in Code Offset encoding

[swaroop-sridhar]

Core Offsets on Thumb/Arm/Arm64 are 2/4/8 byte aligned.
So, GCInfo encoding can save bits by eliminating the last few bits in the offset in NORMALIZE_CODE_OFFSET()

However, one problem with this is in the code-offsets for safepoints, which are encoded with a -1 adjustment:
https://github.com/dotnet/coreclr/blob/master/src/gcinfo/gcinfoencoder.cpp#L1210-L1218
https://github.com/dotnet/coreclr/blob/master/src/vm/gcinfodecoder.cpp#L397-L401
So, currently code offsets are encoded as-is.

So, this work item is to take reduce the number of bits used in the encoding by:

1. Separating the normalization for safepoints (which uses a -1 adjustment) and fully-interruptible code (where no adjustment is used).
2. Have an architecture specific adjustment value (ex: -1 on X86, -2 on ARM, -4 on ARM64) so that we can take advantage of the alignment, and still keep the offset within the safepoint instruction.

category:throughput
theme:gc-info
skill-level:intermediate
cost:medium