bench: foreach vs ReadGroups callback decode benchmark (#156)

[pedrosakuma]

Context

Follow-up to #156 / PR #158 (v1.5.0): we shipped the foreach-style group enumerator with the claim that it's faster and zero-alloc compared to the existing ReadGroups callback API. This PR backs that claim with numbers.

What

Adds GroupForeachVsCallbackBenchmarks comparing three decode strategies on MarketDataData (two simple top-level groups, qualifies for foreach), parameterized over GroupSize ∈ {10, 50, 100}:

1. Callback — ReadGroups with capturing lambdas (mirrors the most common user pattern: closure allocation per call)
2. Foreach — v1.5.0 enumerator
3. Foreach + early break — demonstrates the skip-cost of accessing groups without iterating them

Workload (identical across variants): sum Price.Value + Quantity.Value over every entry, returned to defeat dead-code elimination.

Results

AMD EPYC 7763, .NET 9.0.14, BenchmarkDotNet v0.15.4:

Method	GroupSize	Mean	Ratio	Allocated
Callback	10	124 ns	1.00	152 B
Foreach	10	22 ns	0.17	0 B
Foreach + break	10	3 ns	0.02	0 B
Callback	50	524 ns	1.00	152 B
Foreach	50	211 ns	0.40	0 B
Foreach + break	50	3 ns	0.006	0 B
Callback	100	999 ns	1.00	152 B
Foreach	100	184 ns	0.18	0 B
Foreach + break	100	3 ns	0.003	0 B

Takeaways

• ~5× faster on full iteration at typical group sizes.
• Eliminates the 152 B / call closure allocation (delegate + display class for the captured sum).
• Early break is essentially free because each group property does an O(1) skip; with callbacks you always pay for every entry.
• Validates the docs/perf-tuning-guide recommendation to prefer foreach for simple groups.

Notes

• Existing RepeatingGroupBenchmarks.DecodeWithGroups is left untouched — it does encode+decode in the hot path and isn't directly comparable. New file is focused (decode-only, pre-encoded buffer).
• Schema uses <sbe:message name=\"MarketData\"> which has only simple top-level groups, so it qualifies for the foreach emit path.
• README under benchmarks/ updated with the new benchmark description and reference numbers.