Force 4-byte alignment on the embedded firmware array

[neilrackett]

Problem

COPY_FIRMWARE_TO_RAM_DMA (in rp/src/include/memfunc.h) streams the generated firmware array into ROM_IN_RAM using the RP2040 XIP stream engine. The XIP STREAM_ADDR register truncates its source address to a 32-bit word boundary.

target/atarist/firmware.py generates the image as a bare const uint16_t[], which is only 2-byte aligned. When the linker happens to place it on a 2-mod-4 address, the stream starts reading two bytes early, so the entire cartridge image lands in ROM_IN_RAM shifted by two bytes and the m68k executes garbage.

Because it depends purely on linker layout, the symptom varies per build — bombs, a black screen, or booting straight to the desktop — and appears or vanishes with unrelated code changes, which makes it very hard to diagnose. (The memcpy fallback path in the same macro does a plain uint16_t* copy and is alignment-safe, so only the DMA path is affected.)

Fix

Emit the array with __attribute__((aligned(4))) so the STREAM_ADDR truncation is always a no-op:

const uint16_t __attribute__((aligned(4))) target_firmware[] = { ... };

A one-line change in the generator; the header is regenerated on every build. Verified in a downstream app that was hitting the intermittent crash — with the alignment forced, the cart image copies correctly on every build.