Investigate and polish perf in OpenAPI pipeline

[captainsafia]

We introduced microbenchmarks to the new OpenAPI-generation implementation in preview4. Since then, we've introduced a few more features to the implementation that have perf impacts (more transformers, $ref support, recursive schema transformers). This issue tracks doing a perf analysis on the implementation after all these changes are in to assess impact and trade-offs of making improvements here.

Here's the baseline for numbers run against 28acf62 for our current set of benchmarks.

Method	EndpointCount	Mean	Error	StdDev	Op/s	Gen 0	Gen 1	Gen 2	Allocated
GenerateDocument	10	366.4 μs	5.36 μs	5.74 μs	2,729.52	3.9063	-	-	694 KB
GenerateDocument	100	3,365.4 μs	29.89 μs	27.96 μs	297.14	46.8750	15.6250	-	6,295 KB
GenerateDocument	1000	37,951.0 μs	228.55 μs	202.61 μs	26.35	200.0000	-	-	62,314 KB

Method	ElementCount	Mean	Error	StdDev	Op/s	Gen 0	Gen 1	Gen 2	Allocated
OpenApiSchema_GetHashCode	1	260.1 ns	2.38 ns	2.11 ns	3,844,780.4	-	-	-	-
OpenApiSchema_GetHashCode	10	260.1 ns	3.44 ns	3.05 ns	3,844,440.9	-	-	-	-
OpenApiSchema_GetHashCode	100	262.2 ns	0.95 ns	0.84 ns	3,813,767.2	-	-	-	-

Method	TransformerCount	Mean	Error	StdDev	Op/s	Gen 0	Gen 1	Gen 2	Allocated
ActivatedOperationTransformer	10	2.466 μs	0.0377 μs	0.0477 μs	405,514.3	0.0534	0.0114	-	7 KB
OperationTransformerAsDelegate	10	2.561 μs	0.0195 μs	0.0173 μs	390,494.5	0.0496	0.0114	-	6 KB
ActivatedDocumentTransformer	10	2.462 μs	0.0369 μs	0.0288 μs	406,209.9	0.0534	0.0114	-	7 KB
DocumentTransformerAsDelegate	10	2.399 μs	0.0463 μs	0.0455 μs	416,841.0	0.0496	0.0114	-	6 KB
ActivatedSchemaTransformer	10	31.617 μs	0.4569 μs	0.4274 μs	31,628.4	0.3662	-	-	56 KB
SchemaTransformerAsDelegate	10	30.972 μs	0.6155 μs	0.6320 μs	32,287.4	0.3662	-	-	55 KB
ActivatedOperationTransformer	100	4.082 μs	0.0557 μs	0.0494 μs	244,954.8	0.0687	0.0153	-	9 KB
OperationTransformerAsDelegate	100	3.250 μs	0.0595 μs	0.0497 μs	307,655.7	0.0458	0.0076	-	6 KB
ActivatedDocumentTransformer	100	4.149 μs	0.0428 μs	0.0358 μs	241,048.5	0.0687	0.0153	-	9 KB
DocumentTransformerAsDelegate	100	3.241 μs	0.0258 μs	0.0215 μs	308,574.1	0.0496	0.0114	-	6 KB
ActivatedSchemaTransformer	100	38.633 μs	0.2470 μs	0.2190 μs	25,884.8	0.4883	-	-	66 KB
SchemaTransformerAsDelegate	100	35.940 μs	0.4064 μs	0.3802 μs	27,824.2	0.4883	-	-	62 KB
ActivatedOperationTransformer	1000	34.187 μs	0.4856 μs	0.4543 μs	29,251.0	0.2441	-	-	30 KB
OperationTransformerAsDelegate	1000	11.128 μs	0.0788 μs	0.0658 μs	89,863.8	0.0458	0.0153	-	6 KB
ActivatedDocumentTransformer	1000	36.648 μs	0.6589 μs	0.8332 μs	27,286.5	0.2441	-	-	30 KB
DocumentTransformerAsDelegate	1000	12.282 μs	0.0807 μs	0.0674 μs	81,422.7	0.0458	0.0153	-	6 KB
ActivatedSchemaTransformer	1000	145.059 μs	1.1951 μs	0.9980 μs	6,893.7	1.2207	0.2441	-	164 KB
SchemaTransformerAsDelegate	1000	72.551 μs	0.4466 μs	0.3959 μs	13,783.4	0.7324	-	-	117 KB

[captainsafia]

Parking this in .NET 10 planning since I anticipate any further perf optimizations will ship as part of that release.

One thing I am keen to do is examine whether we are allocating too many OpenApiSchema objects as we construct the model