Add install from source instructions

[mhvk]

I thought I would give the quad dtype a try. It was giving errors (will raise separate issue), so I thought I should perhaps install from source. That doesn't really help, but as it took me a while to figure out how to do it, it seemed a good idea to add instructions to the README. Note that this is super linux-centric, but perhaps better than nothing?

[ngoldbaum]

Fun, not sure why the test is failing, but it's unrelated.

@SwayamInSync would you mind giving this a look? If not I'll try diving in here a bit tomorrow (and see if I can reproduce @mhvk's crash).

[mhvk]

OK, sounds good. Note that I basically started from the github workflow; great that you had that!

[SwayamInSync]

Fun, not sure why the test is failing, but it's unrelated.

@SwayamInSync would you mind giving this a look? If not I'll try diving in here a bit tomorrow (and see if I can reproduce @mhvk's crash).

Sure, I'll give this a go this weekend, a possible hunch is that Sleef got some updates recently (didn't followed them thoroughly) but that can be a possibility of crash.
Thanks @mhvk we will check and fix it :)

[mhvk]

@SwayamInSync - thanks! As for sleef, after I installed sleef on my machine, I did run the tests and they all passed.

[SwayamInSync]

Quick fix and issue

I see, my suspection seems to be correct, it was indeed version changing issues (and a pretty interesting one)
We build this with the 3.6 version of Sleef and now currently ongoing in 3.8

So I changed the version back to 3.6 in my fork and it passed all the tests

git clone -b 3.6 https://github.com/shibatch/sleef.git

Now fun part:

The issue is in the that test too, the test checks for the nan as

assert (QuadPrecision("nan") != QuadPrecision("nan")) == (QuadPrecision("nan") == QuadPrecision("nan"))

A simplified version would be

(nan != nan) == (nan == nan)

As per IEEE-754 this should evaluate to

(nan != nan)  # True
(nan == nan)  # False

# so overall
True == False

So the previous version (3.6) was IEEE incorrect (passing that test) and which when fixed in 3.8 made the error

so the correct test should be

def test_nan_and_inf():
    # NaN should not equal itself
    assert QuadPrecision("nan") != QuadPrecision("nan")
    
    # Test infinity comparisons
    assert QuadPrecision("inf") > QuadPrecision("1e1000")
    assert QuadPrecision("-inf") < QuadPrecision("-1e1000")

@ngoldbaum another reason we should plan a tlfloat backend :)

[mhvk]

I now checked on my laptop and it turns out my installation instructions had some typos and had aftereffects of experimenting with environment variables. I corrected these.

[ngoldbaum]

Thanks @mhvk, let's fix the failing CI in a separate PR.