Read EGI simple-binary event channels in blocks

[bruAristimunha]

What does this implement/fix?

_read_events() walks a simple-binary EGI file one time sample at a time:
for every sample it seeks past the data channels and issues a separate
np.fromfile for that sample's event values. Opening a file therefore costs two
Python-level I/O calls per sample.

Samples are stored as interleaved frames — n_channels data values followed by
n_events event values — so the same bytes can be read a few MB at a time and
the event rows sliced out. Same reads, ~4000x fewer calls.

Numbers

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

file	main	this PR
32 MB, 30800 samples, 256 ch	170.3 ms	10.8 ms	15.7x
shipped test_egi.raw (79 KB)	1.96 ms	1.53 ms	1.28x

Beforehand cProfile put _read_events at 89% of read_raw_egi (0.194 s of
0.219 s, 30821 np.fromfile calls). The cost is linear in recording length, so
it grows with the file — the fixture above is only ~2 minutes of data.

Correctness

• Bit-identical to main: data arrays and annotations compared with
  np.array_equal (the annotations matter here — they are built from exactly
  these event channels).
• Memory stays bounded: the loop reads 4 MiB at a time rather than slurping the
  whole recording, so the seeks are not traded for a large allocation.
• pytest mne/io/egi: 26 passed.

Caveats

No fixture in the testing dataset is long enough to show this, so CI cannot
demonstrate it. The large file was synthesised by tiling the shipped fixture's
data section and patching n_samples in the header; I verified the header
layout matches the file exactly before tiling.

Additional information

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