test(perf): Add remaining Phi-4-mini-instruct perf tests

[venkywonka]

Description

Add Remaining Phi-4-mini-instruct BF16 / FP8 perf results.
The two single-GPU tests that timed out earlier in #4267 now finish well within the CI limits.
Both native BF16 and post-quantised FP8 variants are included.

Performance Summary

backend	in len	out len	quant	concurrency	req throughput (req/s)	output TPS (tok/s)	avg latency (ms)	result
cpp	500	2000	–	1	0.0892	178.31	11 216	PASS
cpp	500	2000	fp8	1	0.1315	263.07	7 602	PASS

Raw metrics

metric	BF16	FP8
Request Throughput (req/s)	0.0892	0.1315
Total Output Throughput (tok/s)	178.31	263.07
Total Token Throughput (tok/s)	222.88	328.84
Average Latency (ms)	11 216.49	7 602.42
P50 Latency (ms)	11 218.60	7 602.73
P90 Latency (ms)	11 222.94	7 614.00
P95 Latency (ms)	11 222.94	7 614.00
P99 Latency (ms)	11 222.94	7 614.00

All runs use max_batch_size = 32, requests = 8, concurrency = 1 on a single H100-80GB GPU.

[Copilot]

Pull Request Overview

Add two remaining performance tests for the phi_4_mini_instruct model in BF16 and FP8 precision, using lower request counts and concurrency to fit CI timeouts.

• Introduced BF16 and FP8 variants with reqs:8 and con:1.
• Added a comment clarifying why request counts were reduced.