Update llama.cpp to v0.8.0 - #45

… split LLM tests into intention-named modules

…num values

…th tests and drop unreachable destructor statuses

…ead variables

…nd parse target-cpu from codegen flags as rustc does

…ders so vendored code no longer warns

…ins, and cap non-causal image tokens to the micro-batch

…e probing them, and test the error paths of sampling, logits, chat sessions and duck typing

…locations as typed errors

…ath, and make test helpers fail fast

…ount is tested, and fold the hf-hub client errors into one

…ve policy

…es and gate the tests crate at full coverage

… coverage

…of guessing it by name

…tus mapping

… test helpers to the support crate and drop debug prints from tests

…ow-0 logits APIs, and fix IMROPE, NUMA init, buffer override ownership and chunk evaluation exceptions

…ory and state operations as llama.cpp does

…orts

…ocab, decode pieces without dropping bytes, and read chat spans as one array

… project and local crates

…rom std::ffi

…e, report exceptions thrown while applying a sampler, and name tests after what they verify

…eturn before evaluating any

… macros at module scope, and name tests after what they verify

…piece with its one allocation, and name every test after what it verifies

…unsafe length setting they guarded

… prefix

…d seeded sampler

… generation ends

…ubstrings

… token

…it from the parser under test

… emitter

…y, reconfigure CMake on change, and drop the redundant backends dir chain

…cargo profiles

…when supported

… byte, and refuse keys llama.cpp cannot hold

…e database when any wrapper-sources file changes

…ve attribute literals as values

…equence ids

…loading from a directory, and write pieces as an outcome

…e instead of guessing gomp

…ree still reports it

…ama.cpp defines it

…es from entries

…reason, provisioning models in their own crate

… two trials with one name