common: sanitize malformed UTF-8 to U+FFFD with retry parse

[renovys]

Problem

Chat template parsing hard-fails when the input contains malformed UTF-8 (e.g. truncated multibyte sequences from upstream tooling or IPC boundaries). The whole parse is rejected even though the rest of the input is usable.

Change

• Add common_utf8_sanitize(): replaces each maximal invalid subpart with a single U+FFFD per the Unicode recommendation, keeps a trailing truncated sequence intact for streaming.
• In the chat parse path, on a failed first parse, sanitize and retry once; parse AST views are then served from the sanitized copy.
• Non-malformed inputs take the same fast path as before (single parse, no copy).

Tests

• test-chat: regression cases for lead-byte + partial continuation runs, overlong/invalid sequences, and mixed valid/invalid payloads. Full suite passes.

Validation

Running in production on a BC-250 (gfx1013, Vulkan) service for ~7h with mixed Korean/security/code workloads — no template parse failures, no template parse failures and no regressions observed.

[renovys]

Rebased cleanly onto the current prism base (adfffbe41, prism-b10743): git cherry-pick applies with no conflicts, and the sanitize+retry path builds and runs on a live BC-250 service. Still a single self-contained commit against prism. Ready for review when convenient.

[renovys]

Rebased onto the current prism (88c4bc60b). While doing so I dropped the unrelated test-backend-ops large-tile cases that had slipped into this PR, so it is now only the UTF-8 sanitize + retry-parse change and its test-chat regression tests (4 files). test-chat passes on this base (CPU build).