323: link.x.in: put most __[se] symbols back into sections r=adamgreig a=jordens

This puts most start/end address symbols back into the sections.

Only `__ebss` and `__edata` are kept outside their sections so that
potential user code with external libraries can inject stuff using
`INSERT AFTER .bss/.data` and profit from the .bss/.data zeroing/loading
mechanism. This also leads to the `__sbss` and `__veneer_base` symbols
having the right section type (bss is B not D in nm).

Also the trust zone start and end address are aligned to 32 bytes as per
the requirements. That section does cost up to 28 byte of padding at the end of FLASH due to
that alignment even if empty. But since flash is typically a multiple of 32 bytes and since the padding is at the end, there is no downside.

The .rodata start is kept free for the linker to allocate it after .text.
This enables users to inject sections between .text and .rodata and
removes the chance to get overlapping address errors. With this the
linker will by default place .rodata after .text as before.

This commit also adds and exposes a few more stable address start/end symbols
(__[se]uninit, __stext, __srodata) that are usefull for debugging and hooking
into.

See
#287 (comment)
for discussion of the issues and description of this compromise solution.

Tested:

* [x] [stm32h7 ITCM](quartiq/stabilizer#322)
* [x] [sgstubs](#323 (comment))
* [x] `INSERT AFTER` for bss/data still works (#323 (comment))

Topics:

* [x] `sgstubs` moved to be the last section in FLASH to minimize the impact of the 32 byte alignment. (Padding flash to 32 byte is considered benign.)
* [ ] `INSERT AFTER` with binutils ld doesn't work. But that's independent of these changes. This is the `sgstubs`-in `memory.x` use case. Currently the `sgstubs` section is kept in `link.x`.

Co-authored-by: Robert Jördens <rj@quartiq.de>