- Controlled by `thread_local` feature
- Does not require nightly features, but using it does
- Intercepts `__pre_init` to copy .tdata into the per-core state
- Implements `__aeabi_read_tp` to returns per-core state for code compiler
  generates when accessing `thread_local` variables
- Needs linker script support to set up the layout and symbols

This is based on
[picolibc](https://github.com/picolibc/picolibc/blob/58d6157cc2135df5043d62c3e89feedc20ffcd57/newlib/libc/picolib/machine/arm/read_tp.S#L71)'s
support for TLS; its the only example I found of it for rp2040.

- for __eabi_read_tp, simplify to just directly load TLS_CORE_[01]
  as needed without the need for an indirection. This only uses r0
  and doesn't touch the stack.
- Write the tdata copy in Rust, being careful to only use raw pointers.
  The generated asm is functionally identical to the hand-written asm.

FIXME: nightly only