Pooled span dicitionary direct pointer

[Scooletz]

This PR changes PooledSpanDictionary to use pointers rather than relative addresses that need to then transformed before accessing. This unfortunately makes the address size twice as big (from uint to byte*). When the Root PageCount is increased to make it big enough to keep the same number of buckets (128kb for the root instead of the earlier 64kb), the new version wins.

Benchmarks

Read_write_small has been modified.

Before pointer

Method	Mean	Error	StdDev	Gen0	Code Size	Allocated
Read_from_big_dict	174,569.2261 ns	3,233.0694 ns	3,024.2151 ns	-	410 B	-
Enumerate_big_dict	69,030.7471 ns	670.5640 ns	627.2460 ns	-	207 B	-
Read_missing_with_hash_collisions	207,001.1198 ns	2,432.6261 ns	2,275.4799 ns	-	416 B	-
Read_missing_with_no_hash_collisions	169,260.6755 ns	1,962.7596 ns	1,835.9666 ns	-	435 B	-
Read_write_small	3,351.7085 ns	22.0201 ns	18.3878 ns	1,364 B	-
Read_colliding_hashes_different_lengths	0.8284 ns	0.0102 ns	0.0095 ns	-	660 B	-

After pointer and 2 times bigger Root

Method	Mean	Error	StdDev	Gen0	Code Size	Allocated
Read_from_big_dict	143,440.3209 ns	2,785.0180 ns	3,420.2525 ns	-	343 B	-
Enumerate_big_dict	41,784.9841 ns	742.0425 ns	694.1071 ns	-	133 B	-
Read_missing_with_hash_collisions	162,168.6432 ns	3,115.8957 ns	3,060.2262 ns	-	349 B	-
Read_missing_with_no_hash_collisions	126,322.7669 ns	2,015.2744 ns	1,885.0889 ns	-	381 B	-
Read_write_small	3,009.3006 ns	35.7414 ns	33.4325 ns	1,301 B	-
Read_colliding_hashes_different_lengths	0.7309 ns	0.0155 ns	0.0145 ns	-	590 B	-

[dipkakwani]

Interesting use of clearing the page in parallel. Makes me wonder if we could make a bulk/batch rent API as well, instead of sequentially renting a new page? Along the same lines, would exposing a batch Set API in PooledSpanDictionary could also be useful?

[github-actions]

Package	Line Rate	Branch Rate	Health
Paprika	81%	78%	➖
Summary	81% (5136 / 6329)	78% (1808 / 2324)	➖

Minimum allowed line rate is 75%