Use ArcSwap for aggregate fn registry

[robert3005]

ArcSwap is faster than a lock for read. These session are mutable but mutations
are rare and retrievals are common

[codspeed-hq]

Merging this PR will not alter performance

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 5 improved benchmarks
❌ 2 regressed benchmarks
✅ 1244 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	WallTime	cuda/bitpacked_u8/unpack/3bw[100M]	353 µs	300.1 µs	+17.63%
⚡	Simulation	encode_varbin[(1000, 2)]	162.7 µs	142.9 µs	+13.87%
⚡	Simulation	encode_varbin[(1000, 32)]	170.1 µs	150 µs	+13.45%
⚡	Simulation	encode_varbin[(1000, 4)]	163.8 µs	143.7 µs	+14.03%
⚡	Simulation	encode_varbin[(1000, 8)]	165.1 µs	145.2 µs	+13.69%
❌	Simulation	new_alp_prim_test_between[f32, 16384]	103.7 µs	118.1 µs	-12.21%
❌	Simulation	null_count_run_end[(10000, 4, 0.01)]	112.2 µs	126.6 µs	-11.41%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing rk/aggregatearcswap (9630c96) with develop (495f30e)}