BitFilter made faster

[Scooletz]

This PR amends BitFilter to make it faster. Tracing shows that up to 8% of the block execution falls into the BitFilter creation. By making it faster, we can improve the time before other endeavors like #417 are taken. This PR:

1. improves Page.OrWith by unrolling the loop and making two vectorized operations per spin that gives ~20% improvement
2. makes or-ing of multiple filters parallel, by applying the operation per range of pages at the specific index. Think running it by column rather then by row. This should help in both cases, where the number of committed sets is low and high.
3. uses SSE prefetch to get the next page before it's ORed
4. modifies the OfN bit filter so that it uses a generic <TSize> parameter that allows to skip the length, mask and shift fields

Overall, BitFilter creation is ~60% faster now (depends on the number of ancestors).

The allocation on the filter comes mostly from Parallel.For that allocates 3.6kb on its own. It looks that we need to pay the cost.

Benchmarks

Page.OrWith

Method	Mean	Error	StdDev	Code Size
OrWith (before)	41.86 ns	0.576 ns	0.510 ns	48 B
OrWith (after)	36.77 ns	0.372 ns	0.330 ns	63 B

BitFiler.OrWith

Method	count	Mean	Error	StdDev	Gen0	Code Size	Allocated
Or_BitMapFilter_OfN_128_Multiple (before)	16	223.0 us	4.13 us	3.86 us	0	376 B	328 B
Or_BitMapFilter_OfN_128_Multiple	16	101.9 us	5.61 us	16.28 us	0.3662	2,370 B	5.56 KB
Or_BitMapFilter_OfN_128_Multiple (before)	32	498.0 us	9.87 us	17.29 us	0	376 B	584 B
Or_BitMapFilter_OfN_128_Multiple	32	108.6 us	2.14 us	3.08 us	0.2441	2,408 B	4.77 KB
Or_BitMapFilter_OfN_128_Multiple (before)	64	1,423.7 us	19.28 us	16.10 us	0	376 B	1097 B
Or_BitMapFilter_OfN_128_Multiple	64	344.2 us	5.55 us	4.92 us	0.4883	2,401 B	6.42 KB

Assembly

The assembly for Page.OrWith

Before

; Paprika.Benchmarks.PageExtensionsBenchmarks.OrWith()
       vzeroupper
       mov       rax,[rcx+8]
       mov       rcx,[rcx+10]
       xor       edx,edx
       nop       dword ptr [rax]
M00_L00:
       vmovups   ymm0,[rax+rdx]
       vpor      ymm0,ymm0,[rcx+rdx]
       vmovups   [rax+rdx],ymm0
       add       rdx,20
       cmp       rdx,1000
       jb        short M00_L00
       vzeroupper
       ret
; Total bytes of code 48

After

; Paprika.Benchmarks.PageExtensionsBenchmarks.OrWith()
       vzeroupper
       mov       rax,[rcx+8]
       mov       rcx,[rcx+10]
       xor       edx,edx
M00_L00:
       vmovups   ymm0,[rax+rdx]
       vpor      ymm0,ymm0,[rcx+rdx]
       vmovups   [rax+rdx],ymm0
       vmovups   ymm0,[rax+rdx+20]
       vpor      ymm0,ymm0,[rcx+rdx+20]
       vmovups   [rax+rdx+20],ymm0
       add       rdx,40
       cmp       rdx,1000
       jb        short M00_L00
       vzeroupper
       ret
; Total bytes of code 63

[github-actions]

Package	Line Rate	Branch Rate	Health
Paprika	83%	81%	➖
Summary	83% (4817 / 5784)	81% (1641 / 2034)	➖

Minimum allowed line rate is 75%