mblock_map on 32-bit targets should be a compact bitset

Currently, on 32-bit targets, mblock_map is defined as a StgWord8 array, with each byte representing a boolean for the corresponding mblock. This is easy to implement but wastes a lot of space. It should be possible to use a compact bitset (implemented by StgWord64 array) to represent mblock_map, with the following pros:

• No waste in space. This can be significant when we want to fine tune the mblock size for certain targets
• getFirstMBlock and getNextMBlock can get a nice speedup, a single memory load with a ctz instruction can check 64 mblock map entries at once.

Potential cons:

• HEAP_ALLOCED() takes more instructions since some bitwise operations are needed to calculate the correct index, and also to mask irrelevant bits. And it's a hot path in the RTS codebase.

Should we change mblock_map implementation? Opening a tracking ticket here and waiting for discussion with other folks familiar with RTS.