Fix PSRAM cache invalidation

[dhalbert]

• Fixes RP2350 PSRAM inconsistency and crashes #9755.

The cache invalidation in #9752 needed to use volatile pointers to point to the cache maintenance entries so they would be set correctly and in the right order.

Note that the code from pico-sdk does use volatile:
https://github.com/earlephilhower/arduino-pico/pull/2551/files
It comes from the discussion here:
https://forums.raspberrypi.com/viewtopic.php?t=378249
That code also uses _compiler_memory_barrier(), which may not be needed, but it helps guarantee ordering.

Besides the example in #9755, this even simpler reproducer caused a hard fault on a Feather RP2350 with PSRAM before the fix:

import gc
gc.mem_free()

After the fix, this works, which is more like the example in #9755 (#9755 doesn't have the first gc.mem_free()):

>>> import gc
>>> gc.mem_free()
8264912
>>> a = bytearray(3500000)
>>> b = bytearray(3500000)
>>> gc.mem_free()
1175376
>>>
    [press reset]
>>> import gc
>>> gc.mem_free()
8264912

@tannewt Also, I don't think the extra 4 bytes is needed just below the PSRAM cache, so I took that out. I looked over the tlsf code, and I think it takes care not to write below the beginning of the pool. It does offset the first block by the size of a header. The first pool created is not passed a pointer one word up in free storage, so I think the second pool does not need to either. But you have spent a lot of time on tlsf, so maybe I am wrong.

@bill88t Could you test if you have a chance?

[bill88t]

Passes every test I threw at it. The only thing I didn't test is for corruption.
Everything in the issue is fixed with this commit, thanks!