Update to gcc 15.3.Rel1 from 15.2.Rel1

[dhalbert]

• Update to gcc 15.3.Rel1 from 15.2.Rel1. Fixes arm-none-eabi-gcc (Arm GNU Toolchain 15.3.Rel1 (Build arm-15.149)) #11283 as part of the update.

I did smoke tests on a few boards.

[dhalbert]

Claude wrote this:

Firmware size, gcc 15.3.Rel1 vs 15.2.Rel1

Compared against the main run for the base commit (1043a70), en_US builds, bytes of flash used. Broadcom and spresense don't report a flash region, so they are not included.

port	boards	median delta	range
raspberrypi	160	-1456	-2056 to -56
mimxrt10xx	16	-880	-904 to -856
nordic	54	-800	-800 to +16
renode	1	-124	-124
analog	2	-76	-80 to -72
stm	24	-60	-912 to +24
silabs	3	-56	-56 to -48
atmel-samd	117	-4	-20 to +20
espressif, zephyr-cp, litex	285	0	0 (control, other toolchains)

Where the savings come from

An nm -S comparison of local 15.2 and 15.3 builds of raspberry_pi_pico_w, feather_nrf52840_express, teensy40 and feather_stm32f405_express shows no symbols added or removed and no data-section changes. The decreases appear to be improved optimization, concentrated in a few unrolled bit-manipulation loops:

• On Cortex-M4 boards, common_hal_bitmaptools_readinto() shrinks from 1172 to 840 bytes.
• On RP2040, common_hal_bitops_bit_transpose() shrinks from 1872 to 1052 bytes.
• The rest is a tail of 10 to 66 byte reductions in jd_decomp(), mp_execute_bytecode(), gc_alloc(), mp_map_lookup(), mpz_mul_inpl() and similar.

This explains the bimodal nordic and stm results: boards without bitmaptools (internal-flash nordic boards, nucleo/discovery/pyboard stm boards) save only a few dozen bytes.

[mikeysklar]

The devcontainer URL 404s: Arm moved 15.3 to gitlab.arm.com. The gcc action already uses the new location, so CI stays green. Suggestion inline.

Built the head with both compilers and ran six ARM boards (RP2040, RP2350, M4, M0, nRF52840, STM32F405). Output identical on every test. Timings within a couple of percent except the RP2 boards, where the interpreter loop moved 25% faster on the RP2040 and 9% slower on the RP2350, which looks like XIP cache placement rather than codegen. pico2_w reproduces #11283 at the first commit and builds at the head.

[mikeysklar]

Devcontainer URL downloads and extracts to the name the symlink expects. Six ARM boards ran identical output on both compilers, and pico2_w reproduces #11283 at the first commit and builds at the head.