olean loader: no-mmap fallback re-reads the entire file after the header probe

In lean_compacted_region_read

1. The 88-byte olean_header is read to validate magic/version/githash
   (readn(fd.get(), &header, sizeof(header)), ~line 482).
2. If LEAN_MMAP is compiled in, the file is mmapped at the saved base_addr.
3. On the fallback — no LEAN_MMAP, or the mapping did not land at
   base_addr — the loader does malloc(size), lseek(fd, 0, SEEK_SET),
   and re-reads the whole file from offset 0 (~lines 558–560), including
   the header bytes it already consumed.

I am working on Lean in WASM64 which is configured with -DMMAP=OFF takes the
fallback for every olean. On Emscripten targets the redundant second read goes
through proxied filesystem emulation; an import trace there measured
38,568 read operations for 19,284 module loads — exactly 2× per module.
Desktop builds also hit the fallback whenever the mmap address check
fails at runtime.

strace -f -e trace=read,lseek lean SomeFile.lean 2>&1 | grep -B1 -A2 '", 88) = 88'

On fallback loads the same fd shows the 88-byte header read followed by
lseek(fd, 0, SEEK_SET) and a full-file read.
I do have a AI assisted patch here if needed https://github.com/FawadHa1der/lean4/commit/2dab7cd11c (guarded by
#ifndef LEAN_MMAP); happy to adapt it into a PR.

Start in src/library/module.cpp at lean_compacted_region_read and inspect the 88-byte olean_header probe and the fallback around lines 558–560. Use the reported strace command, or an equivalent WASM64 import trace, to verify that fallback loads no longer reread the header while still loading the full file.