Move forward update links out of Node struct

[KonaeAkira]

This PR moves all forward update links into a single Vec. Each Node will instead store the begin and end indices for its own links.

This has the following benefits:

• Reduces the size of Node from 112 down to 64 bytes.
• Removes one unsafe usage.
• Improves the repeater_grid benchmark (Add benchmarks for compile time and backend runtime performance #212) by ~3%.
• Removes smallvec dependency.

[BramOtte]

I can confirm this does improve performance.

Iris Mandelbrot benchmark went from about 2'800'000rtps to about 2'950'000rtps or about a 5% performance increase

This was was run at 500'000 ticks at a time on a ryzen r1200 on fedora 43 with no other programs running.

(I think the few slow runs on master were because I did not have the terminal in focus for the entire run, I made sure to keep it in focus for the rest.)

raw runs, master:

DONE in 51.5569016s processing 135000000 ticks, effective rtps: 2618466.118219951
DONE in 52.117169247s processing 135000000 ticks, effective rtps: 2590317.2016152996
DONE in 48.945879831s processing 135000000 ticks, effective rtps: 2758148.397089338
DONE in 47.847170448s processing 135000000 ticks, effective rtps: 2821483.459439198
DONE in 47.789161943s processing 135000000 ticks, effective rtps: 2824908.295337335
DONE in 47.862234891s processing 135000000 ticks, effective rtps: 2820595.409041072

This pull request:

DONE in 45.64615921s processing 135000000 ticks, effective rtps: 2957532.5139387557
DONE in 45.321526445s processing 135000000 ticks, effective rtps: 2978716.971588091
DONE in 46.623534334s processing 135000000 ticks, effective rtps: 2895533.3809078448
DONE in 45.2200781s processing 135000000 ticks, effective rtps: 2985399.5320720156
DONE in 45.242627898s processing 135000000 ticks, effective rtps: 2983911.551388195

[KonaeAkira]

Aligning the Node struct to L1 cache lines further improved the repeater_grid benchmark over the previous commit by ~1.6%.

Current performance gain over master:

repeater_grid           time:   [240.26 µs 241.45 µs 242.92 µs]
                        change: [-6.3477% -5.6568% -4.8247%] (p = 0.00 < 0.05)
                        Performance has improved.

[KonaeAkira]

Running the Mandelbrot program on the BatPU-2 shows a ~17% increase in TPS (no flushing, -io compile flags).

master branch:

Simulation completed in 168.642932388s (1539883 tps)
Simulation completed in 167.267331651s (1549104 tps)
Simulation completed in 167.046934586s (1549104 tps)

This PR:

Simulation completed in 142.849635035s (1821834 tps)
Simulation completed in 143.566302508s (1809094 tps)
Simulation completed in 145.828858343s (1784141 tps)

[StackDoubleFlow]

Thanks! Nice and simple change, and good results