feat: inline instrumentation optimizations

[ndrewh]

Follow-up to #5

from pwnlib.asm import asm
def lib_hook(b):
    b.raw(asm("int3"))

• Calling b.raw for every call to the instrumentation builder is decoding the instruction sequence each time. If we had a way of constructing (thread-global) raw expressions that could be reused, then we could avoid this overhead
• Calls to raw assembly instructions must un-reserve registers, commit the abstract state, and re-reserve registers. This is really inefficient, especially for sequences of raw assembly. At a minimum, we could elide the reservation and commit if there's no other changes (i.e. for hooks that only contain calls to b.raw).

fuck = ctypes.c_long(1337)
def lib_hook(b):
    b.regs.rsi += b.load(ctypes.addressof(fuck))

• I believe dynamorio provides special instruction builders for if fuck has an address reachable from the code cache (e.g. you can do LEA with a 32-bit offset from RIP rather than constructing a 64-bit immediate and loading from that.