fix: use sentinel to distinguish unset config_filename from explicit None

[rahulpar2103]

Problem

When COPY . . is used in a Dockerfile, any local .env file gets baked
into the image. Inside the container, os.path.isfile(".env") returns True
and slowapi loads it, potentially overriding env vars injected by Docker
--env-file or Kubernetes secrets. The only workaround was the undocumented
config_filename="/dev/null" trick.

The deeper issue: config_filename=None (the default) and
config_filename=None (explicitly passed) were indistinguishable, so there
was no clean API surface to say "don't load any file."

Closes #256

Fix

Introduce a module-level _UNSET = object() sentinel. The parameter default
changes from None to _UNSET. The logic inside __init__ is:

• config_filename is _UNSET → existing auto-detect behavior (backward-compatible)
• config_filename is None → skip file loading entirely
• config_filename is a string → load that specific file (unchanged)

Backward compatibility

Anyone who was calling Limiter(key_func=..., config_filename=None) explicitly
was getting auto-detect behavior, identical to passing nothing. That call now
disables file loading. In practice no one passes the default explicitly, but
this is worth noting in the changelog.

Tests added

• test_config_filename_none_disables_dotenv_loading: verifies that
  config_filename=None does not load a .env file present in cwd
• test_config_filename_unset_loads_dotenv_when_present: verifies that
  omitting the argument still auto-loads .env (no regression)
• test_config_filename_unset_no_dotenv_does_not_error: verifies that
  the no-file case does not raise (the original bug scenario)

Docker note

As a complementary mitigation, users should add .env to .dockerignore
to prevent credentials from being baked into images regardless of library behavior.

[laurentS]

@rahulpar2103 thanks for raising this PR. The problem indeed does feel annoying.
I'd be tempted to remove the backwards compatible behaviour here, and only load a config file if explicitly passed. This would both simplify the code in slowapi and follow what Starlette 1.0+ seems to do (and avoid this weird "None means .env" behaviour).
@detrax you raised the original issue, any thoughts on this?