Scale the default CPU thread count with the machine

[felladrin]

Description

Currently, cpuThreads defaults to hardwareConcurrency - 2, which oversubscribes on anything but a small machine. navigator.hardwareConcurrency reports logical processors, so half of it approximates the physical core count on the SMT CPUs most users have, and that is where wllama's throughput peaks.

Measured on a 16-core/32-thread box, in Chromium, with SmolLM2-135M Q4_K_M, a 660-token prompt and the WASM CPU backend (n_gpu_layers: 0), reading wllama's own timings and peak RSS across the browser process tree:

threads	prompt tok/s	generation tok/s	peak RSS
6	59.8	34.9	999 MB
14	81.4	47.5	1008 MB
16	68.3	57.5	1106 MB
22	67.7	34.0	1073 MB
30	53.7	16.5	1535 MB

The peak sits at 16, the physical core count. At 30 threads (what this machine gets today) generation is about 3.5x slower and peak memory is ~39% higher. In a separate run where n_threads matched hardwareConcurrency exactly, leaving no core for the main thread, it degraded much further, to 0.43 tok/s.

The default now scales with the machine rather than using a fixed number, so it lands on the physical core count at any size:

logical processors	before	after
2	1	1
4	2	2
8	6	4
16	14	8
32	30	16

The setting stays user-editable, so anyone who wants more threads can still raise it. The input description already warned that "a value that is too high may cause the app to hang"; this stops the default from being that value.

The thread count now lives in getDefaultCpuThreads() so it can be tested without stubbing navigator.

How to test

1. Run npx vitest run client/modules/settings.test.ts. The new getDefaultCpuThreads tests cover the scaling, the rounding on an odd processor count, and the "never oversubscribe" invariant.
2. Clear the settings key from localStorage, reload, and open Menu => AI Settings. "CPU threads to use" should read half your logical processor count (the value navigator.hardwareConcurrency reports in the console).

Note: the benchmark numbers come from a standalone Playwright harness rather than from this repository, so they are not reproducible from a script here. The regression is measurable in the app by setting cpuThreads to hardwareConcurrency - 2 and then to half, on a machine with many cores, and comparing generation speed.

Only the measurements at 16 and 30 threads were repeated; the other rows are single runs, and the machine was not fully idle.