Add accumulation precision, TPU v6e tests, and SQNR benchmarks for multi-pass FP8 in QWIX

[copybara-service]

Add accumulation precision, TPU v6e tests, and SQNR benchmarks for multi-pass FP8 in QWIX

• Add unit test evaluating multi-pass FP8 across accumulation regimes:
  • FP32 accumulation (unconstrained mantissa)
  • Reference software emulation with output rounded to BF16
  • Internal BF16 accumulator (preferred_element_type=bfloat16), demonstrating exponent alignment truncation of the 2^-4 residual passes
• Add TPU v6e test (multipass_dot_tpu_test) with requires-ghostfish tag to verify execution on Ghostfish hardware.
• Add comprehensive SQNR benchmark summary table across multi-pass FP8 strategies contrasting CPU software accumulation against physical TPU hardware accumulation (512x512, channel-wise / tensor scaling).

Multi-Pass FP8 SQNR Benchmark Table (512x512, Channel-wise / Tensor Scaling)

Strategy	GEMMs	Compute DType	Execution Target	Accumulation Mode	SQNR (dB)	Rel Error (%)
Baseline BF16 (Reference)	1	bfloat16	TPU / CPU	Output Rounded to BF16	55.62 dB	0.17%
Pure FP8 (1-Pass)	1	float8_e4m3fn	CPU Reference	FP32 Accumulation (Rounded to BF16)	28.43 dB	3.79%
Pure FP8 (1-Pass)	1	float8_e4m3fn	Physical TPU	Hardware Accumulator	28.39 dB	3.80%
Triangular Multi-Pass FP8	3	float8_e4m3fn	CPU Reference	FP32 Accumulation (Rounded to BF16)	54.84 dB	0.18%
Triangular Multi-Pass FP8	3	float8_e4m3fn	Physical TPU	Hardware Accumulator	47.48 dB	0.42%
Full Cross Multi-Pass FP8	4	float8_e4m3fn	CPU Reference	FP32 Accumulation (Rounded to BF16)	55.59 dB	0.17%
Full Cross Multi-Pass FP8	4	float8_e4m3fn	Physical TPU	Hardware Accumulator	47.60 dB	0.42%