Add single-threaded njit Xphi kernel for single-process path

[sbwells22]

The single_process=True fallback routes the Xphi computation to compute_Xphi_data_numpy, which carries no numba decorator. That leaves both ways of running multiple trials compromised: running parallel-numba trials concurrently oversubscribes cores (each worker spins up an all-core threadpool), while the single_process alternative avoids the oversubscription only by dropping JIT on the hottest kernel. This PR adds compute_Xphi_data_serial. which is the compute_Xphi_data body with prange -> range, and points the single_process branch in _fit at it. compute_Xphi_data_numpy is left in place as the reference. It's additive and no existing function renamed or changed, no public API touched. The serial kernel is bit-identical to the parallel one with the side benefit that single_process=True now matches single_process=False exactly, where they previously diverged slightly (the numpy fallback derives e_logx and normalizes the softmax differently).

Scope—basically, what this does not fix. This does not address the oversubscription in scHPF_consensus.py. With -j N, consensus launches N concurrent scHPF train subprocesses via subprocess.Popen, and scHPF train calls run_trials with single_process=False where each child spins up its own all-core numba threadpool and you get ~N×cores threads on cores cores, independent of this change. The serial kernel only affects the single_process=True path, which consensus never takes. Fixing that is a separate change, such as capping NUMBA_NUM_THREADS per child, or routing consensus through train-pool/run_trials_pool (which uses single_process=True and would then pick up this kernel).