FP32 coshf_u10 and FP64 cosh_u10 are inaccurate

[hdelassus]

On x64 targets, we found that Sleef_coshf_u10 and Sleef_cosh_u10 have maximum absolute error > 1 ulp.

The entry points for which the errors were reproduced are:

• Sleef_cosh_u10
• Sleef_coshd1_u10
• Sleef_coshd1_u10purec
• Sleef_coshd1_u10purecfma
• Sleef_coshd2_u10
• Sleef_coshd2_u10avx2128
• Sleef_coshd2_u10sse2
• Sleef_coshd2_u10sse4
• Sleef_coshd4_u10avx2
• Sleef_coshd8_u10avx512f
• Sleef_coshf_u10
• Sleef_coshf1_u10
• Sleef_coshf1_u10purec
• Sleef_coshf1_u10purecfma
• Sleef_coshf4_u10
• Sleef_coshf4_u10avx2128
• Sleef_coshf4_u10sse2
• Sleef_coshf4_u10sse4
• Sleef_coshf8_u10avx2
• Sleef_coshf16_u10avx512f

Reproducers are attached. Compile them with:

cc -O3 coshf-test.c -o coshf-test -mavx512f -lm -lsleef -lmpfr
cc -O3 cosh-test.c -o cosh-test -mavx512f -lm -lsleef -lmpfr

Outputs:

$ ./test-coshf
coshf_0(-0x1.634ceep+6) = inf
coshf_1(-0x1.634ceep+6) = inf
coshf_2(-0x1.634ceep+6) = inf
coshf_3(-0x1.634ceep+6) = inf
coshf_4(-0x1.634ceep+6) = inf
coshf_5(-0x1.634ceep+6) = inf
coshf_6(-0x1.634ceep+6) = inf
coshf_7(-0x1.634ceep+6) = inf
coshf_8(-0x1.634ceep+6) = inf
coshf_9(-0x1.634ceep+6) = inf
MPFR faithful results = 0x1.1b925ap+127, 0x1.1b925cp+127
$ ./test-cosh
cosh_0(0x1.62f58b40d30edp+9) = inf
cosh_1(0x1.62f58b40d30edp+9) = inf
cosh_2(0x1.62f58b40d30edp+9) = inf
cosh_3(0x1.62f58b40d30edp+9) = inf
cosh_4(0x1.62f58b40d30edp+9) = inf
cosh_5(0x1.62f58b40d30edp+9) = inf
cosh_6(0x1.62f58b40d30edp+9) = inf
cosh_7(0x1.62f58b40d30edp+9) = inf
cosh_8(0x1.62f58b40d30edp+9) = inf
cosh_9(0x1.62f58b40d30edp+9) = inf
MPFR faithful results = 0x1.252d53b6b0c93p+1023, 0x1.252d53b6b0c94p+1023

These results unduly overflow.