fix: preserve a leading BOM when set_key and unset_key rewrite a file

[MohammedAlkindi]

set_key and unset_key drop a leading UTF-8 BOM from files that have one. #640 made the parser strip the BOM so the first variable parses, but the stripped BOM never reaches the bindings, so the rewrite writes the file back without it:

>>> pathlib.Path(".env").write_bytes(b"\xef\xbb\xbfa=x\n")
>>> set_key(".env", "b", "y")
>>> pathlib.Path(".env").read_bytes().startswith(b"\xef\xbb\xbf")
False

rewrite() now carries the BOM across, so both writers keep it, and only files that already had one change behaviour. The check is encoding-safe: a latin-1 file never decodes to \ufeff, and utf-16 handles its own BOM in the codec.

The new set_key and unset_key tests fail on main on the missing BOM and pass with the fix. The full suite passes, and ruff check and ruff format --check are clean.

Related: #640, #637

[tesdano-crypto]

Windows validation of PR #687

I used an AI-assisted local test script on Windows with Python 3.12.10 to
compare main (a00cb2eed0704cd6d2071b2004c37e95ccc86ee5) with this PR
(b636f75406b9d6e84821f92b2039b0cb5f7bc78c). The synthetic .env files
were stored under a Traditional Chinese directory name containing a space.

The matrix covered 8 operations for each of 4 encoding configurations:
set the first key, set the second key, append a key, unset the first key,
unset the second key, unset a missing key, unset the only key, and set a key
in an empty/BOM-only file. The configurations were UTF-8 with BOM read using
UTF-8, UTF-8 with BOM read using UTF-8-SIG, UTF-8 without BOM, and UTF-16.

All 32 parsed-value checks passed on each revision. For a UTF-8-BOM file
rewritten using the default UTF-8 encoding, main removed the BOM in all
8 cases; this PR preserved it in all 8 cases.

One existing edge case remains: with encoding="utf-8-sig" or
encoding="utf-16", removing the only key leaves a zero-byte file, without
a BOM. This occurs on both revisions, so it is not a regression introduced
by this PR. It may be useful to document or test the intended empty-file
behavior separately. Minimal example:

import codecs
from pathlib import Path
from tempfile import TemporaryDirectory
from dotenv import unset_key

with TemporaryDirectory() as folder:
    path = Path(folder) / "sample.env"
    path.write_text("FIRST=value\n", encoding="utf-8-sig")
    assert path.read_bytes().startswith(codecs.BOM_UTF8)
    unset_key(path, "FIRST", encoding="utf-8-sig")
    print(repr(path.read_bytes()))  # b'' on both revisions

This was a focused local matrix, not a full-suite or CI run. I have not
established downstream breakage from the empty-file behavior.