Serialize build_callable under a global lock to prevent segfaults in threaded usage

[Copilot]

On Julia 1.10, concurrent threads can race through build_callable during first-time callable IR generation, causing segfaults in threaded Particle Gibbs sampling. The existing GlobalMCCache lock only protects setindex!, not the full IR derivation + compilation pipeline.

• Wrap build_callable body in a ReentrantLock so only one thread performs callable IR generation at a time
• Cache lookup is rechecked after acquiring the lock to avoid redundant compilation

const build_callable_lock = ReentrantLock()

function build_callable(sig::Type{<:Tuple})
    # argument validation (no lock needed)
    lock(build_callable_lock)
    try
        # existing body unchanged
    finally
        unlock(build_callable_lock)
    end
end

Fix #226

[github-actions]

Libtask.jl documentation for PR #227 is available at:
https://TuringLang.github.io/Libtask.jl/previews/PR227/