perf: structural cold-build wins to reach 50% faster than PlatformIO #1468

Part of the cold-build perf meta issue.

Removing zccache's miss overhead should get us to roughly parity with PlatformIO. Reaching ≤ 50% of PlatformIO's cold time (≤ ~370 ms against PIO's ~745 ms on ubuntu-24.04) needs structural work, because the raw avr-gcc floor for the 26 TUs is already ~340 ms on a 4-core desktop and is likely higher on CI.

Candidates to evaluate once #1465 data exists (fill in the numbers):

• Critical-path scheduling: start the sketch TU at the same time as the core TUs instead of after them, and start the largest TUs first.
• Overlap archive/link with the compile tail, and skip ar for the core by linking the objects directly.
• Overhead audit: measure fork/exec, env and pipe setup, and the spawn path through running-process. Reuse compiler environment/argv construction.
• Precompiled headers for Arduino.h across core TUs, if the output is bit-identical.
• Content-addressed prebuilt framework core shipped with the package archive. This needs a fairness ruling first: the benchmark's cold definition retains "installed packages/toolchains and fbuild package archives", and PlatformIO keeps its packages too, so decide explicitly whether prebuilt objects count and document it in the benchmark's cold definition.

Rule: no change to the benchmark's cold definition that isn't applied equally to every tool.