The attached raw patch is the current state of the regcache object.  Of 
note:

It is no longer a regbuf but a real regcache object.

It cleans up two of the remaining registers[] cases - stop_registers and 
  dummy frames.

It tries to separate out the pre/post register_read_p cases and for the 
post case, always 1:1 maps [0..NUM_REGS).

I should put more of the register cache description into the `struct 
regcache_desc' object (this object also avoids the need to call on 
gdbarch methods such as MAX_REGISTER_RAW_SIZE).

I suspect RichardE will come up with something for 
{read,write}_register_bytes :-)

enjoy,
Andrew