Gracefully handle a request body that can't be deepcopied

[agu2347]

Fixes #737.

before_record_request() unconditionally called copy.deepcopy(request) before running any filter functions. A request's body can contain something that isn't deepcopy-able, most commonly a raw, open file handle -- e.g. aiohttp file uploads, which are commonly passed directly as part of the request body (data={"file": open(path, "rb")}) rather than pre-read into bytes. deepcopy()ing an open file handle raises TypeError: cannot pickle '...' instances, crashing the entire recording attempt for any request containing one, exactly as described in the issue.

The request object here is VCR's own internal representation (named vcr_request throughout the aiohttp stub, for example), not a reference to the live, actual network request -- so if the deepcopy fails, falling back to using the original object directly (letting filter functions mutate it in place instead of a copy) is safe. It's still far better than crashing and preventing the upload/recording from working at all.

Testing: verified directly: confirmed deepcopy() on a real open file handle reproduces the exact reported TypeError, and confirmed _build_before_record_request()'s returned function crashes on a request whose body contains one, with the original code. With the fix, the same request is handled correctly, and header filtering still takes effect. Also verified the normal case (a request with an ordinary bytes/str body) is completely unaffected: it's still properly deep-copied and isolated from the original request object, exactly as before.

Added a regression test using a real open file handle as part of the request body, going through the actual use_cassette()/filter_request() path. I confirmed the test fails with the original code (raises TypeError) and passes with the fix. Ran the full existing test_vcr.py suite (26 passed: 25 baseline + 1 new), no regressions.