Added bash A profile update scripts and restored ports_arch as their source

[fdesbiens]

Gives the ARMv7-A and ARMv8-A ports the same bash tooling the M profile has, restores ports_arch as their source, and moves the cortex-a CI job off Windows.

Stacked on #591. Merge #590 and #591 first.

Bash equivalents of update.ps1

update.sh sits beside each update.ps1, with the same cores, compilers, copy sets and patches. --copy-example, --copy-validation-test and --copy-regression-test are accepted and ignored, exactly as in the PowerShell version, where they are declared but never consulted, so both take the same command line.

The two were compared by what each reports as drifted, against the last Windows run of the job. They agree exactly on all 63 files, and the bash version reports 12 more.

Those 12 files are a bug in update.ps1

The ARMv7-A patch table writes its two .cproject patterns differently:

('value=`"cortex-a7`"', 'value=`"cortex-$($core_short_lower)`"'),   # backticks, never matches
('Cortex-A7.NoFPU',     'Cortex-$($core_short_upper).NoFPU')        # plain, always worked

The shipped files show precisely that split:

Port	CPU value	FPU string
cortex_a5/ac6	value="cortex-a7"	Cortex-A5.NoFPU
cortex_a9/ac6	value="cortex-a7"	Cortex-A9.NoFPU
a8, a12, a15, a17	same	same

So the AC6 Eclipse example builds for six Cortex-A cores have been naming cortex-a7 as their CPU since the script was written. ARMv8-A uses plain double quotes and works, which is why the bash version matched it perfectly. The bash scripts do what the PowerShell ones intended, so regenerating corrects those twelve files.

Restoring the source

One genuine revert hazard, the same shape as #590:

_tx_thread_smp_time_get had been implemented in the generated ports only. #555 replaced the stub with a real generic timer read in all 24 SMP ports, but never touched ports_arch/ARMv8-A/threadx_smp/common/src/tx_thread_smp_time_get.S, which still held:

    MOV     x0, #0          // FIXME: Get timer

Regenerating would have put that stub back in every ARMv8-A SMP port, replacing a working timer read. The implementation now lives in the source.

The rest is cosmetic and resolves in favour of the source, since these ports are generated: a trailing blank line removed from 38 copies of tx_thread_schedule.S, and comment spacing in one tx_port.h.

Correction to something I said on #591: I attributed the tx_thread_schedule.S drift to the Cortex-A VFP fix and said regeneration would revert it. Measuring it showed otherwise. That fix is already in ports_arch, was never at risk, and the drift in those 38 files is whitespace. The SMP timer above is the real hazard. #591's description has been corrected.

CI

• The cortex-a job moves from windows-latest to ubuntu-24.04 and runs the bash scripts. It was the only job needing a Windows image, and only because of the tooling.
• Every workflow is pinned to ubuntu-24.04. ubuntu-latest already resolves to that image, so nothing changes today; it means a future runner migration is a deliberate commit rather than something that happens underneath the -m32 builds.
• scripts/check_ports.sh now runs the A profile generators too, so one local command covers both families.

A flaw found in the checker itself

While testing, check_ports.sh reported a clean pass on a tree where the generators were absent: their failure was swallowed by the output redirect, so nothing regenerated and nothing drifted. A check that goes green when it did not run is worse than no check. Generator failures are now caught and reported:

FAIL: generator failed: env -C ports_arch/ARMv7-A ./update.sh --port-sets tx ...

Verification

• The bash scripts reproduce the PowerShell drift set exactly, 63 of 63, plus the 12 explained above.
• After the sync, check_ports.sh passes and the generators are idempotent.
• Every removed line across the 51 regenerated files was accounted for: the trailing blank lines, the four .cproject CPU option lines across six cores, the FIXME stub in the source, and one comment's spacing. Nothing else is lost.
• The generator guard was verified by breaking a generator deliberately and confirming the run fails, then restoring it and confirming it passes.