perf: decode text-part transfer encoding only once

[kurok]

What

text/plain and text/html parts were transfer-decoded twice:

1. get_body_raw() → bytes for attachments[].content
2. get_body() → string for text_plain/text_html, which re-runs the same base64/quoted-printable decode before applying the charset.

This decodes the transfer encoding once (get_body_raw()), reuses those bytes for both the attachment content and the text bodies, and applies only the charset step via a new decode_charset() helper — a faithful copy of mailparse's own internal get_body_as_string using the same charset crate (already a transitive dep). Output is byte-identical.

Benchmark

Targeted input — a ~2 MB base64-encoded text/html body (the path this change touches), 200 iterations, same input both builds:

build	min	median	mean
master	8.308 ms	8.841 ms	8.965 ms
this PR	4.390 ms	4.706 ms	4.706 ms

~1.9× faster on base64/quoted-printable-encoded text bodies. The existing tests/benchmark (large_message.eml) is unchanged within noise — that corpus is dominated by base64 attachments, which carry a name param and were only ever decoded once.

Risk

• Behavior-preserving: decode_charset replicates mailparse's charset logic exactly (same crate, same code path). For 7bit/8bit text it reduces to the prior get_as_string(raw); for base64/QP it reduces to get_decoded_as_string() minus the redundant transfer decode.
• Transfer-decode errors still surface (now from the single get_body_raw()?), preserving the existing ParseError contract.
• All 91 correctness tests + RFC corpus pass. cargo clippy --release clean.