Restored ports_arch as the source of truth for the Cortex-M ports

[fdesbiens]

Makes ports_arch/ authoritative again for the Cortex-M ports, so that ports_arch_check passes and the copy scripts stop being a hazard.

Stacked on #589. That branch removes the duplicated function body in the M4/AC6 header; this one regenerates that file among others, so #589 should merge first.

Why this is needed

The Cortex-M ports under ports/ are generated. scripts/copy_armv7_m.sh copies one tx_port.h plus the per tool sources to fifteen M3/M4/M7 targets, scripts/copy_armv8_m.sh does the same for nine M33/M55/M85 targets, and .github/workflows/ports_arch_check.yml runs both and fails if the tree is not reproducible. Those copies are not meant to be edited directly.

They have been, consistently, since #523. On current dev, running the three scripts changes 35 files, so the check is already red. It has not fired because it triggers on pull requests targeting master, while all of this work merged into dev.

Left alone, the next run of those scripts would have silently reverted three separate pieces of work:

Would be reverted	From
The dsb/isb barriers and "memory" clobbers	#523
The IAR assembly header using ; comments instead of C style, which the assembler requires	direct fix to the copies
Moving tx_initialize_low_level.S into example_build for the M33/M55/M85 GNU ports	#514

That last one is worth dwelling on: the arch source still holds the file in gnu/src/ and lists it in the GNU CMakeLists.txt, exactly as it was before #514 moved it. Running the script would put it back and rebuild it into the library.

What this does

Brings the sources up to what the ports actually carry, then regenerates. After this, the scripts are idempotent and ports_arch_check passes.

Two behavioural changes ride along, both deliberate:

1. The Fixed race condition and message loss in Cortex-M GNU ports #523 barriers reach the ac5 and keil variants of M3/M4/M7. Those targets are copy destinations but were outside Fixed race condition and message loss in Cortex-M GNU ports #523's scope, so they never received the fix. They get it now by construction.
2. The barrier after restoring the interrupt posture applies to every tool. Fixed race condition and message loss in Cortex-M GNU ports #523 gave it to the GNU ports only, because GNU was the only toolchain that could be tested. A single shared source cannot express that distinction, and the difference has no technical basis. The identical __asm__ volatile ("... isb 0xF ..." : : : "memory") statement already ships in the AC6 and IAR ports from Fixed race condition and message loss in Cortex-M GNU ports #523, so this is that same statement a second time rather than new exposure to a compiler we cannot run; the delta is a pipeline flush.

A broken header fixed on the way through

ports/cortex_m85/iar/inc/tx_port.h has ended with one more #endif than #if since #523:

UNBALANCED #endif at line 628
final depth = -1

Every other ARMv8-M port is balanced. That header cannot compile as it stands. Regenerating removes the stray directive. This is the second Cortex-M header found this week that a fix left in a non-compiling state, after the M4/AC6 one in #589 — both in files that no CI job builds.

Verification

• The scripts are idempotent after the change: running them again alters nothing, so ports_arch_check passes.
• No port loses anything. Every removed line across the 35 regenerated files was accounted for: eight MSR BASEPRI writes replaced by their clobbered form, five commented out lines replaced by the real barrier, the stray #endif, and #else lines that move rather than disappear.
• Every regenerated header is preprocessor balanced.
• Every Cortex-M port covered by the two scripts now carries the entry barrier.
• No orphaned function body remains in any tx_port.h.

Not covered

ports/cortex_m0/{ac5,ac6,keil}/inc/tx_port.h still lack the #523 barriers. Those are ARMv6-M, have no copy script, and are maintained by hand, so they need a separate decision rather than falling out of this one. The GNU and IAR M0 ports did get the fix.

A follow-up PR will add scripts/check_ports.sh, covering reproducibility, preprocessor balance and code at file scope, so that each of the three defects behind this change is caught mechanically in future.