net: use fd_map_dynamic.c for fib4 hmap

[akhinvasara-jumptrading]

This PR switches the hash map in fib4 from slot_para to map_dynamic to avoid locking overhead. For this use case, occasional inconsistent/torn reads are acceptable (just misroute or drop a couple packets, rarely). They are mitigated by atomic operations on the core next_hop struct. This improves xdp tx perf by 10-20%, depending on the xdp mode / ZC.

[mmcgee-jump]

Can we just switch this stuff to use message passing? Not sure why we need a custom concurrent shared memory structure for this

[github-actions]

Performance Measurements ⏳

Suite	Baseline	New	Change
backtest mainnet-368528500-perf per slot	0.076803 s	0.077098 s	0.384% ✅
backtest mainnet-368528500-perf snapshot load	3.095 s	3.018 s	-2.488% ✅
backtest mainnet-368528500-perf total elapsed	76.803 s	77.098354 s	0.385% ✅
firedancer mem usage with mainnet.toml	997.11 GiB	997.11 GiB	0.000% ✅

[github-actions]

Performance Measurements ⏳

Suite	Baseline	New	Change
backtest mainnet-368528500-perf per slot	0.057988 s	0.050869 s	-12.277% ✅
backtest mainnet-368528500-perf snapshot load	2.576 s	1.669 s	-35.210% ✅
backtest mainnet-368528500-perf total elapsed	57.988409 s	50.868735 s	-12.278% ✅
firedancer mem usage with mainnet.toml	997.11 GiB	997.11 GiB	0.000% ✅

[github-actions]

Performance Measurements ⏳

Suite	Baseline	New	Change
backtest mainnet-368528500-perf per slot	0.053201 s	0.053461 s	0.489% ✅
backtest mainnet-368528500-perf snapshot load	2.546 s	1.795 s	-29.497% ✅
backtest mainnet-368528500-perf total elapsed	53.201319 s	53.461428 s	0.489% ✅
firedancer mem usage with mainnet.toml	997.11 GiB	997.11 GiB	0.000% ✅

[ripatel-fd]

Can we just switch this stuff to use message passing? Not sure why we need a custom concurrent shared memory structure for this

Maintaining replicas of routing tables across different net tiles results in a significantly more complex control plane. The "custom concurrent" part is a single atomic read. Replication of netlink messages is ~hundreds of lines of code and inherently bug-prone due to possibility of dropped updates, and will further decrease performance due to having to poll additional input links.