timer/hpet: many bug fixes and performance improvements + sdfgen bumped to 0.29.0

[dreamliner787-9]

drivers/timer/hpet:

• Greatly reduced timestamp read latency by serving it from the TSC (Time Stamp Counter) rather than the HPET's main counter. As on a real machine, the HPET is not on the same silicon as the processor, so any accesses to it must cross a bus. Thus it is faster if we take the wall clock time from the TSC instead which is local to the processor. On a 3.3GHz XEON W-1250 Comet Lake CPU, previously the driver takes ~25,000 cycles to read the time, now it just takes ~1,800 cycles.
  • This "fastpath" only works if the CPU has an Invariant TSC and expose it's frequency via cpuid. Otherwise, it fallback to reading time from the HPET as before.
  • Prior art: https://docs.redhat.com/en/documentation/red_hat_enterprise_linux_for_real_time/7/html/reference_guide/chap-timestamping
• Renamed from hpet to tsc_hpet.
• Removed unused constants.
• Switched from Message Signalled Interrupt (MSI) to legacy I/O APIC based interrupt routing. As not all HPET implementations support MSI (e.g. QEMU), and each individual machine will have different permitted I/O APIC routing for the HPET. While legacy routing guarantees that comparator 0's IRQ is always wired to I/O APIC pin 2. Which keep things simple across different machines.
• Made timeout calculation more accurate by doing all math in-terms of ticks rather than nanoseconds. To avoid needless conversions between ticks and nanoseconds.
• Fixed potential overflow when converting from large nanosecond timeouts to ticks by using 128-bit wide integer math.
• Stop, reset the main counter and don't start it until we have finished initialising the comparator. Rather than leaving it at whatever stale value that was there before.
• Made all MMIO accesses volatile.

CI:

• Bumped to sdfgen 0.29.0 to fix x86 IRQ trigger rendering in SDF.

tools:

• Added support for "x86_64_generic_vtx" Microkit platform.

examples/*:

• Bumped to sdfgen 0.29.0.

examples/timer:

• Only print messages after setting a timeout. To reduce perceived timer drift on platform that take a long time to print to console, e.g. x86.
• Added support for "x86_64_generic_vtx" Microkit platform.
• Made the timer IRQ on x86 the highest priority IRQ and explained how it is derived.
• Fixed typos in metaprogram.

examples/echo_server:

• Made the timer IRQ on x86 the highest priority IRQ and explained how it is derived.

[midnightveil]

I feel like there's a better way to do this...

Also we have defines for e.g. NS_IN_S, no need to hardcode this constant.

[midnightveil]

Although I guess we've already had the timer-conversion functions discussion, so maybe I should leave this as is for now.

[dreamliner787-9]

The number is for converting femtoseconds in nanosecond, I've added a constant as you suggested.