Mitigate LFU struct tearing using SeqLock

[bitfaster]

Repeat of #593, but for LFU. The lock counter fits inside node padding when the position enum is reduced from a 32 to 16 bit integer (see #590).

This is preferred to the approach outlined in #596, since we avoid allocating a new node and complicating time-based expiry logic. Nodes now contain a 32bit counter which fits inside the padding for the common case (64bit platform, key and value are 64bit objects).

Revisit:

• slight throughput penalty, assuming this is from the enum change. It is possible to eliminate the enum and instead compare to each list ptr.
• code is duplicated for the seq lock itself - see if this can be de-duped.

[coveralls]

coverage: 99.217% (+0.06%) from 99.159%
when pulling 9e85652 on users/alexpeck/tornwrite4
into ab5878b on main.

[bitfaster]

Get bench difference is noise:

Before

Method	Runtime	Mean	StdDev	Ratio
ConcurrentDictionary	.NET 6.0	8.411 ns	0.2609 ns	1.00
ConcurrentLfuBackground	.NET 6.0	29.658 ns	1.3391 ns	3.53
ConcurrentLfuForeround	.NET 6.0	64.598 ns	1.2908 ns	7.69
ConcurrentLfuThreadPool	.NET 6.0	48.558 ns	3.3162 ns	5.78
ConcurrentLfuNull	.NET 6.0	24.832 ns	0.5998 ns	2.96
ConcurrentDictionary	.NET Framework 4.8	15.384 ns	0.5245 ns	1.00
ConcurrentLfuBackground	.NET Framework 4.8	59.298 ns	1.0931 ns	3.86
ConcurrentLfuForeround	.NET Framework 4.8	99.336 ns	2.2780 ns	6.46
ConcurrentLfuThreadPool	.NET Framework 4.8	62.865 ns	0.7662 ns	4.09
ConcurrentLfuNull	.NET Framework 4.8	38.489 ns	0.8632 ns	2.50

After

Method	Runtime	Mean	StdDev	Ratio
ConcurrentDictionary	.NET 6.0	8.307 ns	0.1237 ns	1.00
ConcurrentLfuBackground	.NET 6.0	29.964 ns	1.8776 ns	3.61
ConcurrentLfuForeround	.NET 6.0	64.969 ns	1.1785 ns	7.82
ConcurrentLfuThreadPool	.NET 6.0	47.788 ns	1.7784 ns	5.75
ConcurrentLfuNull	.NET 6.0	26.173 ns	0.9501 ns	3.15
ConcurrentDictionary	.NET Framework 4.8	15.353 ns	0.4578 ns	1.00
ConcurrentLfuBackground	.NET Framework 4.8	55.862 ns	1.2768 ns	3.64
ConcurrentLfuForeround	.NET Framework 4.8	100.165 ns	1.7840 ns	6.53
ConcurrentLfuThreadPool	.NET Framework 4.8	75.223 ns	1.7748 ns	4.90
ConcurrentLfuNull	.NET Framework 4.8	38.881 ns	0.6930 ns	2.53

[bitfaster]

Read+write throughput is regressed slightly, but eviction is not: