Fix full-byte substitution matrix indexing

[kimmanbo]

Summary

This is a extra hardening over #1132 .

This PR makes substitution-matrix lookup tables cover the complete unsigned-byte range and prevents signed char values from being used as negative array indexes.

It also includes two small portability fixes discovered while validating the change.

Problem

The ASCII substitution matrix and aa2num mappings used UCHAR_MAX as an exclusive upper bound. Because UCHAR_MAX is itself a valid byte value, this omitted 0xFF from the allocated and initialized ranges.

Additionally, several scoring paths converted sequence characters directly to int. On platforms where char is signed, bytes above 0x7F could become negative indexes and cause out-of-bounds access.

Character classification and conversion functions such as isalpha and toupper also require their input to be representable as unsigned char or EOF.

Changes

• Allocate and initialize aa2num for every value from 0 through UCHAR_MAX.
• Expand the fast substitution matrix to cover the complete unsigned-byte range.
• Update nucleotide, amino-acid, and reduced-alphabet mapping loops to include UCHAR_MAX.
• Convert raw sequence bytes to unsigned char before matrix indexing.
• Pass character-classification inputs through unsigned char.
• Preserve the existing behavior of mapping unknown byte values to X.
• Add the missing <sstream> include in the search workflow.
• Mark the file descriptor unused on platforms without posix_fadvise.

Validation

• Built the mmseqs target successfully in the debug configuration.
• Validated the complete byte range, including 0x80 and 0xFF, under AddressSanitizer using a temporary regression harness.
• Ran the existing reduced-matrix test under AddressSanitizer.
• Verified the patch with git diff --check.

A permanent regression test is intentionally deferred until it can be integrated with an appropriate test framework.