Coalesce more impls into the new BufferProvider framework

[sffc]

Progress on #1107
See #837
See #1246

I wanted to do this as a separate PR so I could measure impact on code size. In the main code size example (fixeddecimal_tiny), making this change results in a very small code size increase; the maintainability and flexibility of the new approach is well worth the tradeoff, in my opinion.

[Manishearth]

I love this, and this means that multiple providers over FFI is no longer a codesize hazard!

[sffc]

We can compare code size based on CI now 😁

On main:

-rwxr-xr-x 1 runner docker 32261040 Dec 10 10:26 optim0.elf
-rwxr-xr-x 1 runner docker   755984 Dec 10 10:26 optim1.elf
-rwxr-xr-x 1 runner docker  1962408 Dec 10 10:28 optim2.elf
-rwxr-xr-x 1 runner docker  2914296 Dec 10 10:28 optim3.elf
-rwxr-xr-x 1 runner docker    97760 Dec 10 10:29 optim4.elf
-rwxr-xr-x 1 runner docker    58768 Dec 10 10:29 optim5.elf

On this PR:

-rwxr-xr-x 1 runner docker 32179056 Dec 10 10:29 optim0.elf
-rwxr-xr-x 1 runner docker   977168 Dec 10 10:29 optim1.elf
-rwxr-xr-x 1 runner docker  1568432 Dec 10 10:30 optim2.elf
-rwxr-xr-x 1 runner docker  2412928 Dec 10 10:30 optim3.elf
-rwxr-xr-x 1 runner docker    99848 Dec 10 10:31 optim4.elf
-rwxr-xr-x 1 runner docker    60472 Dec 10 10:31 optim5.elf

I consider the change small enough to be in the noise. Looking at the IR, the main size difference is Rust thinking it needs to call some additional destructors given the one extra layer of indirection.

I'm eager to check this in, and consider a follow-up to improve code size for static data (perhaps in coordination with #1290).