Cache parsed message templates

[nblumhardt]

This integrates the well-tested template cache from serilog/serilog. In general, this will reduce allocations and speed up logging when events use templates properly (16 GB T470p):

Before

Method	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
Native	370.5 ns	7.067 ns	6.265 ns	1.00	0.00	0.0887	-	-	280 B
NoId	1,017.8 ns	10.906 ns	9.107 ns	2.75	0.04	0.2785	-	-	880 B
LowNumbered	1,318.5 ns	11.772 ns	10.435 ns	3.56	0.06	0.3529	-	-	1112 B
HighNumbered	1,363.2 ns	20.672 ns	18.325 ns	3.68	0.07	0.3777	-	-	1192 B

After

Method	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
Native	362.2 ns	3.419 ns	3.031 ns	1.00	0.00	0.0887	-	-	280 B
NoId	604.3 ns	11.431 ns	10.133 ns	1.67	0.03	0.1421	-	-	448 B
LowNumbered	843.9 ns	12.429 ns	11.018 ns	2.33	0.04	0.2155	-	-	680 B
HighNumbered	878.3 ns	17.330 ns	16.211 ns	2.43	0.03	0.2413	-	-	760 B

Performance might regress in apps that (mis)use string concatenation to construct log messages, but the benefits for carefully-written have been found to outweigh this downside in our experiences with Serilog to date.

[skomis-mm]

Oops, sorry, didn't know it auto close referenced PRs too 😄

[skomis-mm]

LGTM 👍