Decode Persyst and Nihon Kohden data in cache-sized blocks

[bruAristimunha]

What does this implement/fix?

Two readers materialise the entire requested range before calibrating it, so
every intermediate is as large as the request and none of it stays in cache.

Persyst had a hand-rolled copy of _read_segments_file() plus one extra
full-size temporary: np.reshape(record, (n_chs, -1), order="F").astype(np.float32).
_mult_cal_one() already casts, so that .astype was a whole-array copy for
nothing. The method is now a call to the shared helper — a net deletion.

Nihon Kohden reads the whole request and then builds four full-size
temporaries over it (+ 0x8000, .astype(int16), * cal to float64, +=/*=).
Same work, now a block at a time.

Numbers

Median of 5 interleaved process pairs against a pristine main worktree:

case	main	this PR
Persyst, 107 MB	88.7 ms	39.3 ms	2.26x
Nihon, 106 MB	267.6 ms	147.7 ms	1.81x
Persyst shipped fixture	0.12 ms	0.12 ms	1.05x
Nihon shipped fixture	0.362 ms	0.364 ms	0.995x

Persyst's gain splits cleanly: 1.32x from dropping the redundant .astype
alone (shared helper at its 100 MB default), the rest from the smaller block.

The Nihon shipped-fixture figure is a median of 8 runs x 200 reps — at 0.36 ms a
single-shot comparison is pure noise (an early 7-rep run read 0.86x). That file
is 5800 samples and fits in one 1 MiB block, so it keeps its current code path
exactly.

Why the two constants differ

The optimum tracks time points per block, not bytes. Nihon has 25 channels and
builds four temporaries per block; its sweep improves monotonically down to
~256 KiB before falling off a cliff at 64 KiB, so 1 MiB is chosen for margin.
Persyst sits at 16 MiB. This is the same concern I raised on #14246 — a
max_block_samples= parameter would express it better than a per-reader byte
budget, and I am happy to go that way instead.

One behaviour change worth flagging

Dropping Persyst's .astype(np.float32) changes results for a 32-bit .dat
whose raw counts exceed 2**24, where float32 was silently rounding them. That
is a precision fix rather than a regression, and 16-bit files (and the shipped
32-bit fixture, whose counts are well inside the range) are unaffected.

Correctness

• Bit-identical to main across every readable Persyst and Nihon Kohden
  fixture, including the synthesised large ones (np.array_equal, data and
  annotations).
• pytest mne/io/persyst mne/io/nihon mne/_fiff: 164 passed.

Large files were synthesised by tiling — Persyst derives its sample count from
the .dat file size, and Nihon's single data block was tiled with
record_duration patched. No shipped fixture is big enough to show the effect,
so CI cannot demonstrate it.

Additional information

AI disclosure: I directed the work and reviewed and tested every change; Claude
Code (Claude Opus 5) profiled the readers, ran the A/B measurements and made the
edits under my direction.

[larsoner]

A speedup plus fewer (executed) lines, love it, thanks @bruAristimunha !