fix(codec): eliminate uninitialized memory in encode_item on unwind

[nathanielford]

Motivation

Fixes #2720.

In tonic/src/codec/encode.rs, encode_item previously advanced the buffer's logical length by 5 bytes using unsafe { buf.advance_mut(HEADER_SIZE); } before invoking the user-provided Encoder::encode. If the encoder panicked or unwound, buf was left with 5 uninitialized bytes exposed to callers catching unwinds.

Solution

• Replace unsafe { buf.advance_mut(HEADER_SIZE); } with safe zero-filled initialization via buf.put_slice(&[0u8; HEADER_SIZE]), eliminating the unsafe block entirely.
• Writing 5 zero bytes into the cache line that is immediately overwritten by finish_encoding incurs minimal performance impact (a single store instruction).
• The alternative RAII drop guard approach was deliberately avoided because it preserves unsafe code and could resurface the soundness vulnerability if std::mem::forget (or a similar leak) were ever introduced.
• Add unit test encode_item_exception_safety_on_panic verifying that unwinding panics leave only initialized zero bytes in the buffer.

Test Plan

• Ran cargo test -p tonic codec::encode.
• Ran cargo fmt --all --check.
• Ran cargo clippy.