More immediate-friendly object header words

[sorear]

(all estimates assume #754 #774 #789 have already happened)

• ag32, arm7: These are already quite immediate friendly, most header words fit in a 15-16 bit immediate field.
• x64: Most of our header words need 34ish bits, but the most convenient instructions for loading them have a 32 bit field. If we bumped size_len up to 40, we'd be able to use 32-bit instructions for most header words, saving 5 bytes each (~200kb total) with little impact on anything. (There are ~40k header word immediates in the basis-less bootstrapped compiler.)
• mips, arm8: The most useful immediate fields are 15-16 bits. size_len=52 would allow a single immediate to be used for most header words; this leaves 8 bits for the tag, which seems low to me but it's the same amount we provide on 32 bit targets. Would save ~160kb for arm8, ~320kb for mips.
• riscv64 (Major stretch goal): size_len=40 would save an instruction for most header words, but not bytes (a 2+2+4 becomes a 4+4, so save 160kb in the rare non-RVC case.) Pushing size_len to 56 would only allow for datatypes with 16 constructors, which seems unacceptably small, but we can get around this by rearranging bits in the header word: either have extra tag bits at the top (requiring two shifts to extract the length, and making pattern matching datatypes with a large number of constructors more complicated), or just leave the 12 low-order bits of header words unused and use lui to initialize them (20 bit field).
• riscv32 (possible future work): Leaving 12 bits unused is not going to work with only 32 to start with, so either the tag field or the size field would need to be broken.