But what does it mean by "bounded storage"?
Typically, the bounded storage means that an adversary cannot store all of the bits that were exchanged in the protocol. For example, in key exchange, the eavesdropper cannot store the whole exchange of the honest parties. In section 1.1 of the first paper you cite, it mentions that the bounds are usually as follows. For a security parameter $n$, the honest participants exchange $O(n^2)$ bits and need $O(n)$ bits of storage to perform the protocol while the adversary needs $\Omega(n^2)$ bits of storage to break the scheme.
Should it assume the length of the message is always too long for any computer on the earth to store?
The bounded storage assumption should hold against anyone who has access to the transmission in the scheme. So if you're communicating over local area networks for example, the memory bound would need to apply to every machine within range. On the other hand, if you're performing key exchange over the internet, the memory bound would need to be quite large.
So can I conclude that the cryptographic scheme that replies on bounded storage model is not so efficient(if it is not very inefficient)?
It depends on the metric of efficiency and in which context you apply the schemes. They are memory efficient, but do require a large transmission.
Since it should always send a very long message. And I think it is not compatible with blockchain(hard to imagine a very long message is stored on the blockchain). Am I right?
If the message is stored on a public blockchain, then by definition you are not in the bounded storage model since many computers all over the world have a copy of the bits exchange, so they can all break the scheme.
I should also mention that there is a quantum variant of the bounded storage model, where it is assumed that the adversary has limited quantum storage. This model is better justified than the classical model since storing quantum information is actually very hard and one of the main challenges in building a scalable quantum computer. Moreover, sending and measuring qubits is actually quite easier technologically. You can buy commercial boxes that achieve this. And the schemes are conceived so that honest participants need no quantum memory whatsoever. The main downside is that these boxes need to be linked by a quantum channel and sending quantum signals over long distances is also difficult. If you're interested to learn more, I suggest starting with Christian Schaffner's thesis.