fix: preserve task response header bytes

[Ricky-7-Yan]

What does this PR address?

Task-result response headers are serialized from ASGI's Latin-1 byte representation, but deserialization currently uses UTF-8 through the default str.encode(). A header such as b"caf\xe9" therefore returns as b"caf\xc3\xa9" after a round trip.

This change uses the serde's existing HEADERS_ENCODING in the response path, matching request deserialization and restoring byte-for-byte symmetry. It adds a focused regression test for a non-ASCII Latin-1 header value.

This revisits the voluntarily closed #5693 from @feiiiiii5 after revalidating the behavior on current main.

Verification:

• direct async red/green round-trip reproduction — failed before the change and passed after it
• ruff check src/_bentoml_impl/tasks/serde.py tests/unit/bentoml_io/test_tasks_serde.py
• ruff format --check src/_bentoml_impl/tasks/serde.py tests/unit/bentoml_io/test_tasks_serde.py

AI assistance was used for implementation and test preparation; the reported commands were run locally.

Before submitting:

• Does the Pull Request follow Conventional Commits specification naming?
• Does the code follow BentoML's code style, pre-commit run -a script has passed? Focused Ruff lint and format checks passed; the full pre-commit suite was not run.
• Did you read through contribution guidelines and follow development guidelines?
• Did your changes require updates to the documentation? No user-facing documentation change is needed for this serialization bug.
• Did you write tests to cover your changes?