Add SVE-accelerated Vec::retain_mut for aarch64

[ywxt]

The PR adds SVE support for specified width types(8, 16, 32 and 64 bits) in Vec::retain_mut. Due to pointer provenance being stripped by intrinsics) here it has to use inline asm instead of sve intrinsics.

1. retain half (ns/iter)

Elements	u32 SVE	u32 scalar	Change	u64 SVE	u64 scalar	Change
4	10.32	12.36	/	10.25	11.83	/
8	12.93	15.08	/	13.40	14.74	/
16	18.66	19.46	/	19.08	18.89	/
32	31.84	31.12	/	31.78	31.26	/
64	32.99	59.17	-44.2%	59.86	59.44	/
1,000	471.51	811.71	-41.9%	820.98	827.16	/
10,000	4,660	7,990	-41.7%	5,836	8,242	-29.2%
100,000	46,414	79,608	-41.7%	57,561	82,861	-30.5%

2. retain whole

Elements	u32 SVE	u32 scalar	Change	u64 SVE	u64 scalar	Change
4	3.46	3.11	/	3.45	3.45	/
8	4.90	4.49	/	4.83	6.22	/
16	8.44	8.14	/	8.44	11.74	/
32	15.83	15.51	/	15.83	22.79	/
64	21.57	30.66	-29.6%	30.72	44.89	/
1,000	358.53	483.33	-25.8%	469.18	696.43	/
10,000	3,127	4,745	-34.1%	5,082	6,912	-26.5%
100,000	32,509	49,501	-34.3%	49,484	79,749	-38.0%

r? @Amanieu

[Amanieu]

• These should be pub(super).
• Why is there such a large difference between the constant for LEN_4 and LEN_8? In fact these constants don't even make sense, I would have excepted smaller sizes to have a larger minimum length.
• There should be a separate constant for size 1, for consistency. Or alternatively have a single constant for minimum byte size.
• These should use SIZE instead of LEN. LEN refers to the vector length rather than the element size.

View changes since the review

[ywxt]

Half-Retain:

Len	u8 SVE	u8 Scalar	Ratio	u16 SVE	u16 Scalar	Ratio	u32 SVE	u32 Scalar	Ratio	u64 SVE	u64 Scalar	Ratio
4	20.57	7.95	0.39x	20.00	6.91	0.35x	26.87	10.19	0.38x	26.64	10.29	0.39x
8	21.98	9.88	0.45x	22.60	10.20	0.45x	30.25	12.30	0.41x	30.36	12.98	0.43x
16	22.71	15.55	0.68x	19.70	15.54	0.79x	32.51	18.61	0.57x	31.24	18.67	0.60x
32	21.50	26.58	1.24x	23.62	27.19	1.15x	29.62	31.14	1.05x	39.00	31.40	0.81x
64	27.71	53.85	1.94x	28.36	55.55	1.96x	38.52	58.14	1.51x	54.43	60.41	1.11x
128	43.50	99.20	2.28x	50.17	100.11	2.00x	57.06	105.99	1.86x	97.07	113.08	1.16x
256	78.77	189.70	2.41x	75.25	191.59	2.55x	154.53	199.86	1.29x	163.49	230.88	1.41x
512	129.10	368.45	2.85x	158.73	375.65	2.37x	300.51	403.08	1.34x	309.97	463.14	1.49x
1,000	304.11	721.12	2.37x	329.30	748.53	2.27x	580.19	783.04	1.35x	599.89	907.82	1.51x
10,000	2,510.14	7,177.15	2.86x	3,105.05	7,456.09	2.40x	5,755.80	7,925.75	1.38x	6,493.38	9,123.48	1.41x
100,000	25,468.15	71,774.77	2.82x	27,775.71	74,522.75	2.68x	57,397.32	79,699.40	1.39x	65,362.17	92,278.75	1.41x

Whole-Retain:

Len	u8 SVE	u8 Scalar	Ratio	u16 SVE	u16 Scalar	Ratio	u32 SVE	u32 Scalar	Ratio	u64 SVE	u64 Scalar	Ratio
4	18.81	2.76	0.15x	15.43	2.76	0.18x	15.56	2.76	0.18x	15.75	2.76	0.18x
8	17.10	4.48	0.26x	16.92	4.33	0.26x	17.26	4.32	0.25x	19.72	4.32	0.22x
16	17.92	8.08	0.45x	17.85	8.29	0.46x	19.45	8.08	0.42x	19.28	8.13	0.42x
32	21.52	15.97	0.74x	20.28	15.82	0.78x	20.16	15.55	0.77x	27.13	15.49	0.57x
64	21.79	31.00	1.42x	21.23	30.70	1.45x	25.98	30.63	1.18x	36.78	31.15	0.85x
128	35.84	60.51	1.69x	38.42	60.62	1.58x	41.78	60.92	1.46x	71.65	61.05	0.85x
256	68.08	130.79	1.92x	71.29	130.79	1.83x	81.51	130.80	1.60x	135.93	130.78	0.96x
512	124.97	250.00	2.00x	146.74	250.01	1.70x	203.80	250.03	1.23x	323.93	249.99	0.77x
1,000	247.01	469.68	1.90x	267.12	469.35	1.76x	317.79	469.36	1.48x	622.36	469.58	0.75x
10,000	2,485.53	4,611.28	1.86x	4,060.09	4,646.64	1.14x	3,752.13	4,611.98	1.23x	6,266.50	4,818.49	0.77x
100,000	23,721.95	48,340.95	2.04x	27,156.28	46,246.36	1.70x	34,872.15	46,536.10	1.33x	54,606.28	47,030.48	0.86x

This is the result on my aarch64 machine(without SVE2.1). LEN_8 should be 64.

I'm not sure if it has smaller sizes for u16 and u8 because of lack of SVE2.1

[ywxt]

And for Whole-Retain, long sequences of u64 have regression here. I'm not sure if it's related to my environment or there is no a fast-path like scalar.

[ywxt]

@rustbot ready

[Amanieu]

I was thinking of waiting for #158312 to land first so that we can reduce the amount of inline asm needed, but I don't think that is necessary.

View changes since this review

[rustbot]

This PR was rebased onto a different main commit. Here's a range-diff highlighting what actually changed.

Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers.