In regards to mGBA, it's already fairly mature and has ironed out a lot of subtle bugs over time. There is potentially space for another GBA core that's cycle-accurate (mGBA is
cycle-count accurate). Higan is the modern existing cycle-accurate emulator that I'm aware of. However, cycle-accurate may not offer any benefits; it's possible it would lead to more accuracy, but mGBA may be accurate enough to resolve identically in all practical cases.
As far as new features go, it's mostly a matter of getting the features into BizHawk, as mGBA already has most of the features you mention. The one feature I'm also hoping to be able to TAS GBA with link-cable. This currently presents multiple issues; mGBA's implementation is not entirely feature-complete (it doesn't support Single-Pak currently), and it may not yet be deterministic to the degree required for TASing. I'm hopeful that these issues will eventually be resolved in mGBA.
As far as sub-frame inputs go, it's probably not relevant for most GBA games, since most games only poll once per frame anyway. Sub-frame resets could potentially be useful for SaveRAM manipulation, but probably would only work in very specific games. endrift said mGBA already supports sub-frame inputs, so it's really just a matter of connecting it to BizHawk and then having BizHawk present it a usable way. Sub-frame in itself an
ongoing project complete mess that BizHawk is not yet designed to support for BizHawk in general, from some recent discussion I've seen in IRC.
AFAICT, for mGBA, it's really a matter of having someone familiar with the internal working to be able to interface with BizHawk. I'm hoping to become more familiar; my immediate motivation is a longstanding and overlooked bug related to SaveRAM not loading when loading savestates.
