[kernel] Ensure all words aligned on word boundaries

[ghaerr]

Pursuant to the discussion with @Mellvik in calculating 8088 instruction execution speed and on devising a precision timer in Mellvik/TLVC#71.

The biggest find here is that the kernel interrupt stack (which is used whenever a hardware interrupt takes place) was sometimes (only when CONFIG_TRACE set) allocated at an odd address - which on an 8086 CPU increases stack read/write (e.g. PUSH/POP) by 4 cycles each instruction!! Not a super big deal since production versions have CONFIG_TRACE off, but still quite a potential slowdown on real hardware during development.

Data produced by the C compiler is automatically aligned, but all ASM data should use .p2align 1 for word variables at the start of the .data or .bss section. .comm variables have alignment specified in their directive separately.