This commit addresses a security vulnerability where `Vec::set_len` was used to "initialize" buffers before data was actually written to them. This pattern can expose uninitialized memory if a panic occurs during the subsequent write operations, as safe code (like `Drop` implementations) could observe the uninitialized data.

The fix involves:
- Using `Vec::spare_capacity_mut()` to obtain a slice of `MaybeUninit<u8>` for compression/decompression operations.
- Deferring the `Vec::set_len()` call until after the operation has successfully completed.
- Ensuring buffers are `clear()`ed before use to maintain consistent state.
- Replacing unsafe `set_len` with safe `resize(..., 0)` for internal buffers that require initialization.

Affected areas:
- `src/stream.rs`: `DeflateEncoder::flush_buffer`
- `src/compress/mod.rs`: Parallel compression for large inputs and `dp_path` buffer.
- `src/batch.rs`: `BatchDecompressor::decompress_batch`

This commit fixes a security vulnerability where `Vec::set_len` was used to expose uninitialized memory during compression/decompression, and resolves a CI failure caused by Adler-32 checksum overflows.

Security Fixes:
- Replaced premature `Vec::set_len` with `spare_capacity_mut()` in `DeflateEncoder` (streaming), `Compressor` (parallel), and `BatchDecompressor`.
- Length is now updated only after data has been successfully written.
- Replaced unsafe `set_len` with safe `resize` for the `dp_path` internal buffer.

Adler-32 Fixes:
- Reduced `BLOCK_SIZE` and `MAX_CHUNK_LEN` to 4032 to prevent `u32` accumulator overflows in SIMD paths.
- Fixed mathematical errors in scalar tail processing and intermediate accumulations.
- Ensured proper `DIVISOR` modulo operations are applied consistently across all architecture-specific implementations (SSE2, AVX2, AVX-512, NEON, DotProd).
- Unified output formatting to `(s2 % DIVISOR) << 16 | (s1 % DIVISOR)`.

Verified with targeted tests in the sandbox.